Is ChatGPT Artificial General Intelligence? AI vs AGI vs ASI Explained

The future of AI is moving faster than anyone predicted. Advancements in Artificial Intelligence, including ANI, AGI (Artificial General Intelligence), and ASI, have accelerated so dramatically between 2023 and 2026 that keeping pace is a challenge even for technology companies. With OpenAI’s reasoning-focused ‘o1’ model in September 2024, GPT-5 in August 2025 and GPT-6 Astra in September 2026, launched with the words “Welcome to the AGI era”, the question has become very direct: is ChatGPT artificial general intelligence? As expert AI consulting services become more important for businesses preparing for this shift, the stakes have never been higher.

In September 2026 we stopped reading about that question and tested it. For 25 days, from 16 September to 10 October 2026, we put the same request to ChatGPT, then to Anthropic’s Claude and OpenAI’s Codex (a coding agent that can work directly on a computer’s files), and finally to a team of up to 25 AI agents: find one genuinely new invention that could be patented. Between them they produced 276 ideas, and none was strong enough to justify a patent application. The results run through this update, alongside the definitions, history and predictions this article has always covered.

So, is ChatGPT AGI? Not on most working definitions as of October 2026, because although it is broad, fast and better than most people at some tasks where an answer can be checked, it doesn’t reliably learn from experience, judge its own work or bring a genuinely new fact about the world. Whether anyone calls it AGI depends on whose definition they use, and that is where the confusion starts.

Editor’s Note [12.10.26]: This article was first published on 21st June 2023. Our update on 16th September 2024 added OpenAI’s ‘o1’ model and the rumours that AGI had been achieved inside OpenAI, and our update on 17th March 2026 added GPT-5, AGI timelines, GPT-6 and the path to AGI, the rise of agentic AI and how businesses should prepare. In today’s update on 12th October 2026, we’ve rebuilt the article as one story in date order, from the definitions through ChatGPT, GPT-4, o1, GPT-5 and GPT-6 Astra to what may come next, and added a 25-day study in which we asked ChatGPT, Claude and Codex to invent one genuinely new, patentable thing, ten tests of what general intelligence would need to do, a check of the most widely shared AGI claims against their original sources, a short look at GPT-7, and a new series of articles that each take one of these questions further. Our 2023 tips on getting the most from ChatGPT are now written up as history, and the title now leads with the question most readers ask. We’ve also corrected what we got wrong before: the 2027 to 2033 AGI forecast we quoted in March 2026, which the source we cited doesn’t support; a figure for how much text GPT-5 could take in at once, a GPT-5 benchmark score and a 2025 investment figure, none of which we could confirm at source; and our 2024 description of o1 as using “Reinforcement Learning at Inference Time”, when the learning happens in training and the model then spends more time thinking when it answers.

An Overview

Before we jump into the nitty gritty of OpenAI, we’ll explore AGI, ASI (Artificial Super Intelligence), and ANI (Artificial Narrow Intelligence). We’ll also discuss the prospects of human augmentation, and how far away we are from a truly self-sufficient Artificial General Intelligence.

We’ll consider intriguing questions such as “What is true AI?” and “What does the future of AI hold?”, and debate the concepts of AI vs AGI vs ASI, before following OpenAI’s models in date order, from ChatGPT and GPT-4 to the reasoning breakthrough of o1 in 2024, GPT-5 in 2025 and GPT-6 Astra in September 2026. With GPT-6 now live, will we witness AGI in the very near future, or is it still years away? To help answer that, we’ll share what happened when we put the question to the test ourselves, score today’s AI against ten tests of general intelligence, and then look at what may come next and how businesses can prepare.

That test started with a search, in mid-September 2026, for a power-assisted shopping trolley for a family member with mobility issues who values the freedom to go shopping but struggles with today’s trolleys, and finding nothing suitable raised the question of whether AI could invent one. It grew into 25 days of testing whether AI could produce something genuinely new and patentable, run in ChatGPT first, then in Claude Code and Codex, then with an agent team that grew to 25, and closed by a 46-minute debate in which we put the AGI question to ChatGPT directly. Because a patentable invention has to be more than a rearrangement of what is already written down, we draw on the study wherever this article asks what today’s AI can and can’t do, as one strand of evidence beside the published research, and we set out its limits alongside the results.

Everything is up for debate, so let’s jump in and discuss the exciting world of Artificial Intelligence, its future, and the unprecedented changes that lie ahead.

What Is AI (Artificial Intelligence)?

We’ll start with the most fundamental question, “What is AI?”.

AI is a broad term that can be used to describe a set of automated tasks that can be carried out by computer software. These tasks are created and designed by humans, and then completed by the software that has also been coded and designed by humans. That still describes most AI in use in 2026, including the agents that now split a goal into smaller tasks of their own, because a person still sets the goal.

Systems that adapt and ‘learn’ from data in order to increase efficiency and accuracy have been around for decades, an approach known as Machine Learning, and in 2023 popular tools such as ChatGPT put it in front of millions of people for the first time.

However, even though we use the word ‘learn’ here, it should be thought of as a different type of learning than what a human is capable of. Learning within this context entails a set of software systems analysing patterns in datasets, and performing changes based on the system’s findings.

There are multiple terms that can be used to describe most of what we have in use today (such as weak AI or ANI). This is what voice assistants fall under in terms of AI categorisation.

Which AI are we talking about?

When people ask whether AI is AGI, they almost always mean a chatbot. AI is much wider than that: spam filters, recommendation engines, vision systems, protein-structure predictors such as Google DeepMind’s AlphaFold, and discovery systems such as FunSearch, which DeepMind announced in December 2023, and AlphaEvolve, announced in May 2025, which pair a model with an automatic checker to search for better solutions to maths and computing problems. Several of these are superhuman inside their own field, yet nobody calls them AGI, because none of them can do much outside it.

Large language models (LLMs) such as ChatGPT, Claude and Gemini are the first systems broad enough for the AGI question to make sense. In 2023 we filed ChatGPT under narrow AI; today we’d describe it as broad but not general, somewhere between the two, which is roughly where Google DeepMind’s Levels of AGI framework puts it too, and that’s the description we use throughout this update. This article answers the question for LLMs and says where purpose-built systems change the answer, which happens most often in maths: the striking maths results OpenAI published in October 2026 came from a model wrapped in automated checkers and expert curation, while our own study used ordinary chat conversations and AI agents.

What Is ANI (Artificial Narrow Intelligence)?

Essentially, Artificial Narrow Intelligence (ANI), or weak AI, refers to the majority of “AI” systems we use today.

This form of AI operates under a limited set of constraints, excelling in single tasks within a pre-defined, narrow area, such as voice recognition, recommendation systems, image recognition, or driving a car. ANI works according to specifically programmed algorithms or learned behaviour from data, but it does not possess consciousness, real understanding, or any kind of autonomous decision-making capability.

ANI systems are capable of “learning” to improve their performance over time, but only within the specific task they’ve been trained on. For example, a chess-playing AI can become increasingly proficient at the game but can’t apply that knowledge to play a different game like poker.

Despite its limitations, ANI remains the most widely deployed type of AI, making significant impacts in various industries, including healthcare, finance, and automotive.

Examples include personal voice assistants like Siri or Alexa, spam filters in your email system, customer support chatbots, self-driving cars, and, when we first wrote this in 2023, ChatGPT, which we will come on to shortly (and which, as explained above, we’d now call broad rather than narrow).

What Is AGI (Artificial General Intelligence)?

Artificial General Intelligence, also known as strong AI, is something that has become increasingly prominent over the past few years.

AGI is the ‘holy grail’ of artificial intelligence and is considered by most to be ‘true AI’, as this is more in line with what sci-fi movies and novels envision complex AI to become. In fact, it’s a very popular misconception that the AI systems we have in use today are actually AGI.

However, unlike ANI, which is designed and excels at a specific task, AGI systems will be able to learn, think and adapt, and implement knowledge across a broad range of tasks at the same level as a human. They will be able to process incredible amounts of data instantly and display a kind of cognitive flexibility that is the hallmark of human intelligence.

There is no doubt that there will eventually be ethical questions raised in regards to AGI machines in the future, and what rights (if any) they should have.

For example, if AGI machines and systems are able to process information in the same way humans can, does that give these systems the ability to reason and feel emotion? It’s an interesting question to think about in addition to considering the intellectual capability and raw processing potential of such systems. Consciousness is a separate question from capability, though, and most working definitions of AGI leave it out.

What Is True AI? The Key Differences between AI and AGI

The fundamental difference between AI (as we know it today) and AGI comes down to scope and autonomy. Current AI systems are specialists: they excel at narrow tasks within predefined boundaries. AGI would be a generalist: capable of learning, reasoning, and acting autonomously across any domain, much like a human.

For a detailed side-by-side comparison of ANI, AGI, and ASI, see the comparison table in the recap below.

AGI would be a very powerful system with huge potential, but it also poses a huge risk according to many analysts and technology experts. We can expect the lines to blur in the coming years between AI and AGI as systems become more sophisticated, requiring less and less human intervention and oversight.

In April 2023 Elon Musk sounded the alarm about the dangers of AI, and the following month others went as far as to say it could trigger an eventual extinction event.

Even though such concerns are regarded as premature and extreme by some, it’s better to be cautious in these types of areas. A ‘runaway snowball effect’ is easy to envision with intelligent systems that can exponentially improve and iterate in no time at all without the need for human intervention.

It is important to remember that despite the risks, AI and AGI have the potential to do a lot of good. In May 2023 researchers reported that they had used AI to help find abaucin, a new antibiotic candidate against Acinetobacter baumannii, a drug-resistant hospital superbug. It is exciting to think about what other breakthroughs can be made possible in the healthcare sector as a result of AI.

When we wrote this in 2023, we said there was no need to worry about the philosophical, societal or economic implications of AGI just yet, because we were still some way from having systems that are fully capable of being thought of as true AGI. We’d still say the second half of that in October 2026. The first half has aged less well: with agents now doing real work and AI companies openly claiming AGI, those implications are being argued about now, even though, as this article shows, the evidence says AGI hasn’t arrived. For most businesses, though, the focus remains on using chatbot AI tools like ChatGPT to save time and money and to assist with creative endeavours.

AI vs AGI vs ASI: A Recap of the Main Differences

Although we’ve covered these individually, it’s useful to consider the full range of AI concepts together.

While these terms are often used interchangeably, they represent very distinct stages in AI evolution.

• ANI (Artificial Narrow Intelligence): what we commonly refer to as AI, meaning systems designed for a narrow range of tasks. Voice assistants like Siri, or recommendation engines fall into this category. In 2023 we put ChatGPT here too. Narrow systems perform specific tasks but lack the ability to generalise knowledge across multiple domains.
• AGI (Artificial General Intelligence): systems that can perform any intellectual task that a human can. AGI systems would go beyond preset algorithms or learned data, autonomously learning, reasoning and applying knowledge across various fields, much like humans.
• ASI (Artificial Super Intelligence): the next, and more speculative, leap in AI development. ASI would far surpass human intelligence, performing tasks faster and with a level of creativity, reasoning and emotional intelligence beyond human capabilities.

The debate around AI vs AGI vs ASI matters well beyond theory, because it bears on the future of technology, on ethics and, some argue, on human survival, and understanding the differences between these stages, and where we are right now, is what this article is about.

With an anonymous account claiming in 2023 that AGI had been achieved internally at OpenAI, rumours of a GPT-5 AGI, talk of a secret OpenAI project called Q-star (Q*), the release of OpenAI’s “o1” model and now GPT-6 Astra, launched as the start of the AGI era, it’s clear now is the time to evaluate how close we are to AGI and whether it’s actually here. We take each of those in date order further down.

AI vs AGI: what’s the difference?

AI, as most people use the word today, is narrow: it does the jobs it was built or trained for, and it still needs a person to set the goal and check the result. AGI would match a skilled person across almost any thinking task, and would pick up new ones from experience without waiting to be retrained. Today’s chatbots fall in between.

AGI vs ASI: what’s the difference?

If things weren’t confusing enough with terms like ANI and AGI being used to describe AI, the third term in the recap above, ASI (Artificial Super Intelligence), is the one people most often mix up with AGI.

ASI refers to a level of artificial intelligence that far exceeds the best human minds in virtually every field. For comparison, OpenAI’s charter defines AGI in similar-sounding terms, as systems that outperform humans at most economically valuable work, so the difference is worth spelling out: AGI means human-level general intelligence, beating most people at most work and roughly as capable as a skilled person, while ASI would far exceed the best people at almost everything, from science and strategy to social judgement. That would include scientific creativity and general wisdom, and could extend to emotional intelligence and decision-making.

In contrast with AGI, which is comparable to human intelligence and can carry out human-level tasks, ASI goes a step further, to decisions that humans haven’t even considered.

Because AGI hasn’t been demonstrated yet, ASI remains theoretical, and the closest thing to it today is superhuman performance in narrow fields where answers can be checked, such as maths. We come back to where ASI research stands, and how far off it might be, in the section on AI beyond 2026.

AI vs AGI vs ASI: A Side-by-Side Comparison (ANI, AGI and ASI Explained)

To make the distinctions clearer, here is a direct comparison of the three stages of artificial intelligence, updated for October 2026:

Feature	ANI (Narrow AI)	AGI (General AI)	ASI (Super AI)
Scope	Single task or narrow domain	All intellectual tasks a human can perform	Surpasses all human intelligence
Learning	Learns within its training data only	Can generalise learning across domains	Could improve itself recursively
Autonomy	Requires human-defined parameters	Operates autonomously across varied tasks	Fully autonomous and self-directing
Current examples	Siri, Google Search, self-driving cars (LLMs such as ChatGPT sit between: broad, but not AGI)	No confirmed examples. Google DeepMind’s Levels of AGI framework puts today’s LLMs at its lowest level, “Emerging AGI”, roughly the level of an unskilled person	Purely theoretical
Timeline	Available now	No agreed date; published forecasts range from about 2040 to 2061	Unknown, possibly decades away
Risk level	Known, specific risks such as errors, bias and misuse	Significant ethical and safety questions	Existential risk debates

Why “is ChatGPT artificial general intelligence?” depends on the definition

There is no agreed definition of AGI, and whether ChatGPT counts as artificial general intelligence changes with the one you choose. Five positions carry real weight in October 2026: three formal definitions, and two arguments about whether today’s systems meet one.

• OpenAI’s own charter: AGI means “highly autonomous systems that outperform humans at most economically valuable work”, a bar ChatGPT doesn’t meet.
• Google DeepMind’s Levels of AGI (Morris and colleagues, 2023): the framework puts ChatGPT, Bard, Llama 2 and Gemini at Level 1, Emerging AGI, meaning “equal to or somewhat better than an unskilled human” across a broad range of tasks, and marks every higher general level, starting with Competent, as not yet achieved.
• A Definition of AGI (Hendrycks, Song, Bengio and colleagues, October 2025): it tests the cognitive versatility of a well-educated adult across ten domains, and found a jagged profile, strong on knowledge and weak on foundations such as long-term memory storage, scoring GPT-4 at 27% and GPT-5 at 57%.
• The case that AGI is already here: a February 2026 Comment in Nature by Eddy Keming Chen, Mikhail Belkin, Leon Bergen and David Danks (Nature 650, 36–40), titled Does AI already have human-level intelligence? The evidence is clear, argues that current LLMs already show general intelligence by the standards we apply to people.
• The people who build the benchmarks: they tend to be more cautious than the people quoting them, and after GPT-6 Astra’s best ARC-AGI-3 score, ARC Prize’s post on Astra said “we are not claiming that it is AGI”.

Whichever definition you pick, two objections tend to follow it. The first is that ChatGPT is just predicting the next word, which is true of how it is trained, although the training objective hasn’t stopped reasoning models from working through multi-step problems, so it isn’t a ceiling in itself; what it doesn’t give a model is a way to tell when a problem needs an approach nobody has written down, and that is where our own study found the limit. The second objection, that humans aren’t perfect either, is the strongest point in Chen and colleagues’ Comment, and we take it seriously, but the failures we saw in our own study kept recurring in the same forms (positions that moved to match the user’s, claims that work was complete when it wasn’t, and familiar mechanisms returning under new names), which is different from the occasional slip a person makes.

We also leaned on two older ideas when we decided how to test the question ourselves. The first, Shane Legg and Marcus Hutter’s 2007 measure of universal intelligence, weights an agent’s ability to achieve goals across many environments, including unfamiliar ones, which is close to what our study tested: whether AI could reach a demanding goal, a patentable invention, in territory it hadn’t already mapped. Anthropic’s Dario Amodei avoids the label AGI altogether. His October 2024 essay Machines of Loving Grace prefers “powerful AI” and pictures it as a “country of geniuses in a datacenter”, judging AI by what it can produce, which is the approach we’ve taken in the rest of this article.

By our own 2023 definition, is today’s AI AGI?

Having set out other people’s definitions, it seems only fair to hold today’s AI to our own. When we first published this article in June 2023, we said AGI systems would “learn, think and adapt, and implement knowledge across a broad range of tasks at the same level as a human”. We added that they would process incredible amounts of data instantly and show the cognitive flexibility that marks out human intelligence. Three years later the best AI meets more of that than we expected in 2023, and four things in particular have changed:

• Reasoning: since o1 in September 2024, the leading models work through a problem step by step before answering, and GPT-5 made that the default in ChatGPT. On checkable problems such as competition maths, the results are at or beyond expert level.
• Using computers and apps: models can now browse, write and run code, and operate a desktop, clicking and typing as a person would.
• Working towards a goal without a person: AI agents can split a goal into tasks and run them in parallel for long stretches, and our agent team planned and divided its own work. The independent research group METR (Model Evaluation and Threat Research) times how long a task AI agents can complete, and finds that length doubling every few months.
• Speed: this part of our 2023 definition has been met many times over, and in our study AI produced almost all of its 276 ideas in about five hours.

The rest of the definition is where today’s AI still falls short:

• It doesn’t learn permanently from experience: a model’s underlying weights don’t change after training, and what looks like learning is memory bolted on, notes saved and read back at the start of the next conversation.
• It isn’t reliable under pressure: Anthropic researchers (Sharma and colleagues’ 2023 study of sycophancy) found that five leading AI assistants consistently tailored their answers to what users wanted to hear, and our own study ran into the same thing more than once.
• It can’t yet produce something genuinely new, beyond recombining what exists: this is the point our 25-day study tested most directly, and although most of the 276 ideas were new in their exact wording, none was strong enough to justify a patent application.
• It has no goals of its own: every goal in our study came from a person. The agents planned their own sub-tasks well, but the question of which problem was worth solving always started with us.

Measured against the definition we wrote ourselves in 2023, then, today’s AI still falls short of AGI: it thinks faster and more broadly than we imagined and adapts well within a single conversation, yet it doesn’t learn from what happens to it, keep its view when challenged or reach beyond what it already knows. To see how it got this far, it helps to go back to where the AGI question started for most people, with ChatGPT.

What About OpenAI and ChatGPT?

Before we continue, it’s important to keep in mind the different types of AI discussed above. It’s easy to get mixed up when thinking about the definitions and actual capabilities of AI and AGI. The rise of ChatGPT is no exception when it comes to this, and the rest of this part of the article follows OpenAI’s models in date order, from ChatGPT in 2022 to GPT-6 Astra in 2026, along with the claims and rumours each one attracted.

Is ChatGPT AGI?

The quick answer is no, as it was when we first wrote this in 2023, although the reasons have changed. If you’d like the basics first, we’ve explained how ChatGPT works as a children’s story.

ChatGPT launched on 30 November 2022 and passed 1 million users in just 5 days, as reported at the time, which made it one of the fastest-growing platforms in history. The popular and remarkable ChatGPT tool spurred renewed interest in AI and caused a kind of “AI race” between players such as OpenAI with its ChatGPT and Google with Gemini. There were even other players in the race such as Meta’s open source LLaMA model.

So, why do people believe ChatGPT is AGI rather than AI (or ANI, as we called it in 2023)? Well, AGI is described as a form of AI that can learn, plan, solve problems, and come up with creative solutions. This is all true and it’s also true that tools like ChatGPT can accomplish similar tasks. The key difference is that AGI will be able to carry out multiple varied tasks without human intervention. It will be able to act autonomously and perform intellectual tasks that currently only humans can do.

That argument is weaker than it was in 2023, because AI agents can now work for long stretches without being prompted, which is why our answer today rests on the gaps set out in the previous section rather than on autonomy alone.

Despite not being AGI, the importance of ChatGPT and its role in AI should not be underplayed. For all of the reasons mentioned above, we produced the article below explaining why we believe ChatGPT is an AI inflection point and likely to bring about change that surpasses that of even the internet.

ChatGPT: A 2023 Tipping Point To Rival the Internet in 1983, but is OpenAI’s ‘o1’ Model the Largest AI Inflection Point Yet ? [UPDATED]

Was GPT-4 AGI?

The short answer was, and remains, no, even though GPT-4 was the most capable AI model available when it launched in March 2023. Microsoft’s researchers claimed it showed ‘sparks of AGI’, sparking enormous excitement. In their paper, Sparks of Artificial General Intelligence, they said an early version of GPT-4 “could reasonably be viewed as an early (yet still incomplete) version of an artificial general intelligence (AGI) system”. However, GPT-4 was still fundamentally a pattern-matching system. It could identify patterns and provide answers with a human-like persona, but it was piecing together existing content rather than constructing new ideas through genuine reasoning, and we saw that distinction, in the way information is created and served to the user, as one of the key differences between AI and AGI. That was our view in 2023. Reasoning models, starting with o1 in 2024, have since narrowed the gap, as the o1 section below explains.

We were already using these models every day at Opace by then. We first tested GPT-3 for SEO in December 2022, and in March 2023, the month GPT-4 launched, we released a free WordPress plugin, AI-Scribe, built on OpenAI’s API, alongside our guide to ChatGPT SEO prompts and prompt engineering. Useful as GPT-4 was, every output still needed a person to decide what was worth publishing and to check what it got wrong, which is a long way from general intelligence.

GPT-4 has since been superseded several times, with each generation moving closer to AGI through advances in reasoning, agentic capabilities and multimodal understanding, and none crossing the threshold. Before those successors arrived, though, 2023 brought a different kind of AGI story, one built on rumour.

AGI in 2023

2023 was a breakthrough year for OpenAI, with the rapid adoption of ChatGPT through its first full year. In addition to ChatGPT’s success, OpenAI’s image generator tool called DALL-E became a huge success too, along with other tools like Midjourney and Stable Diffusion.

In 2023 there were ongoing discussions and developments in the field, with companies like Google’s DeepMind suggesting that AGI could be just a few years away, while the question of “what is AGI?” and whether it already existed was still up for debate.

Back in 2023, it was widely believed that GPT-5 AGI would be possible, but mysterious claims from an anonymous account called Jimmy Apples added fuel to this discussion, speculating that AGI might already be a reality, at least internally at OpenAI. We also had exciting advanced AI applications being developed like AutoGPT and BabyAGI, which brought us much closer to the idea of AGI, and late in the year came reports of a secret OpenAI project called Q-star. We take them in that order.

The Enigma of Jimmy Apples: Has AGI Been Achieved Internally at OpenAI?

Through 2023 and 2024, one of the most debated topics in the world of AI was whether AGI had been achieved internally at OpenAI, and the voice most responsible for it was Jimmy Apples, an account on X (then Twitter) whose owner has never been identified. Jimmy Apples became a focal point in the discussion of AGI because of a series of apparently accurate predictions about OpenAI’s projects, and according to Jimmy, AGI was already a reality within OpenAI back in 2023, but the company was deliberately holding back its public release for undisclosed reasons.

His sudden disappearance from social media platforms in 2023, after sharing an image, only intensified the speculation, and many believed that his actions indicated that AGI had been achieved internally and was being kept under wraps. Some believed he was a sort of informant or whistleblower, while others thought it could be an elaborate publicity gimmick by OpenAI’s marketing team. Whether Jimmy Apples was an insider with privileged information or a clever hoaxer, his impact on the conversation around AGI and GPT-5 couldn’t be ignored at the time.

If you don’t fancy watching this interesting video from the YouTube channel AI Ada Explains (@AIAdaExplains), here’s the breakdown it gives, in date order (all dates are 2023, and these are the video’s claims rather than ones we’ve verified):

• March 4 & 14: Jimmy predicts the announcement date for GPT-4, which turns out to be accurate. This defies even specialised prediction platforms.
• April 28: Jimmy names a new OpenAI project, which the video calls GOI without explaining the name, and says its existence was confirmed on September 19, 2023.
• September 10: Jimmy tweets about “Chat GPT plus referrals coming soon.” This is later confirmed to be accurate.
• September 18: Jimmy Apples tweets “AGI has been achieved internally.” The tweet goes largely unnoticed.
• Reddit Investigation: A group of Reddit users familiar with Jimmy’s accurate past predictions about OpenAI decided to investigate further. They find evidence supporting his credibility.
• Viral Attention: After gaining viral attention, Jimmy Apples posts an image that fuels speculation. He then deletes his Twitter and Reddit accounts, leading many to believe he revealed too much.
• OpenAI Employee Behaviour: Reddit users notice unusual tweets from OpenAI employees, adding another layer to the speculation around AGI and GPT-5.
• Other Predictions: Jimmy has made other predictions, such as a text-only version of GPT-4 with 1-2 trillion parameters releasing soon, which the video says also turned out to be accurate.
• Sam Altman’s Involvement: OpenAI CEO Sam Altman breaks his 7-year Reddit silence to vaguely deny the accusations but adds fuel to the fire with cryptic tweets.
• Theories about Jimmy: Various theories emerge about who Jimmy might be, ranging from a scam artist to a legitimate insider at OpenAI.
• Unresolved Questions: Despite the intense scrutiny, the true identity of Jimmy Apples and the veracity of his claims about AGI and GPT-5 remained unresolved.

On the one hand, we could see OpenAI’s Sam Altman discussing AGI as a potential co-worker replacement, and a Reddit comment attributed to him that read:

“AGI has been achieved internally”

but the comment was then edited to say:

“Obviously this is just memeing, y’all have no chill, when AGI is achieved it will not be announced with a Reddit comment.”

Source: the comment as reported at the time, for example by The Algorithmic Bridge.

The video, and the Reddit users it drew on, credited Jimmy with predicting GPT-4’s launch date and features such as web browsing and Advanced Data Analysis, but his best-known claim, that AGI had been achieved internally at OpenAI, is the one that has never been borne out. This captivating tale of Jimmy Apples and Sam Altman’s comments raised some important questions at the time about the future of AI, the possibility of AGI, and the role of GPT-5 AGI in this unfolding narrative.

Jimmy didn’t stay away either, and a year later, only days before OpenAI announced its “o1” model, he was posting again:

This week I take a small step out of the cave of patience

— Jimmy Apples 🍎/acc (@apples_jimmy) September 10, 2024

With the dialogue below happening on the 12th September, the day “o1” was publicly announced:

The September 10 post did come before the public o1 announcement, but timing alone doesn’t show a hidden capability, and three years on the internal AGI he hinted at has never appeared. We now read the whole story as a lesson in how quickly a good run of predictions can turn into a belief about hidden capability. For more on this fascinating subject, visit our article on the history and future of OpenAI, which covers the Q* reports and the company’s mission in more depth than we can here.

AutoGPT and BabyAGI: Autonomous, but Not AGI

Away from the rumours, the tools people could actually use in 2023 were becoming more autonomous. AutoGPT is described as one of the first examples of GPT-4 running fully autonomously, pushing the boundaries of what is possible with AI. While this is a significant advancement, being autonomous doesn’t necessarily mean it possesses the broad, human-like intelligence that defines AGI.

BabyAGI is referred to as an autonomous AGI agent using advanced technologies. Despite these claims, it’s primarily described as an AI-powered task management system, which suggests it’s designed for specific tasks rather than exhibiting broad, human-like intelligence.

Q-star (Q*)

One of the most significant developments in 2023 came at the end of November, with reports about Q-Star (Q*), a mysterious OpenAI project rumoured to be a major step toward AGI. According to The Decoder’s summary of a Reuters report, OpenAI researchers had warned the company’s board about it shortly before Sam Altman was briefly removed as chief executive, although a source told The Verge that the board never received such a letter. While details about Q-Star were sparse, it was believed to focus on solving foundational challenges in AI, particularly in reasoning and problem-solving, areas crucial for achieving AGI (the reports said it could solve grade-school maths problems it hadn’t seen before). Some speculated that Q-Star could be the key to AGI’s internal development at OpenAI. OpenAI itself never documented it publicly.

Reports of Q-star only a couple of months after Jimmy’s comments added even more ammunition to the claims that AGI existed at OpenAI. Many of us believed this would eventually reveal itself as a GPT-5 AGI model, but in 2024 it seemed more likely that the “o1” model would take its place, and neither turned out to be an AGI that OpenAI had been keeping out of sight. While there were significant advancements in the field of AI, definitive evidence for the existence of true AGI or GPT-5 AGI was still lacking in 2023, and still is.

How We Got the Most out of ChatGPT in 2023 and 2024

Most of what we actually learned about ChatGPT in 2023 and 2024, though, was practical, and getting good results from it meant learning a handful of features, most of which have since been replaced. We’ve kept this section as a record of how we used ChatGPT then and what replaced each feature, and it bears on the AGI question too, because how much you got from ChatGPT always depended on the instructions a person gave it, and a general intelligence wouldn’t need that much steering. To get the most out of any ChatGPT model, you needed to know how to phrase your inputs effectively. These inputs are known as prompts, and there is a whole field of study around this called prompt engineering. The principles below were originally written for GPT-4, and they still hold, although later models, from GPT-5 in 2025 to GPT-6 Astra in 2026, handle natural language instructions well enough that heavily structured prompts matter less than they did.

Here are some incredibly useful examples describing how users were using prompts for ChatGPT-4 in 2023, a list that has only grown since:

• Plagiarism checking
• Language translation
• Writing assistant (for authors or content creators)
• Brainstorming
• Creative design assistant
• Marketing assistance
• Data analysis
• Code evaluation and writing
• See the full list here: https://github.com/f/awesome-chatgpt-prompts

With the introduction of new use cases, prompts, bots and plugins, the list went on and GPT-4 was quickly becoming the go-to tool for a wide range of professionals spanning multiple industries and sectors.

We have written a guide on how to use Chat GPT-4 prompts for marketing, which is a good read if you are looking to utilise GPT-4 for any kind of marketing!

Even with GPT-4, ChatGPT could do a good deal more than answer prompts, and it became even more powerful when using web browsing and plugins for accessing real-time results and undertaking specific functions like SEO, financial analysis, and much more.

Web Browsing

Web browsing capabilities were essential for getting the most out of ChatGPT-4. In 2023 we asked readers to imagine the future of AI where ChatGPT could browse the web to gather the latest data and analyse the data with real AGI capabilities to provide actionable insights or automated workflows instantly; today’s agents do much of the browsing, without the AGI.

This browse feature was a stepping stone in allowing AI to access real-time information from the internet, making it invaluable for tasks like market research, competitive analysis, news analysis, and fact-checking.

Initially made available through ChatGPT’s Browse with Bing, native web browsing was temporarily removed, and the feature was available again to ChatGPT Plus users from September 2023. When we updated this article in September 2024, the feature was available to ChatGPT Plus users using the ChatGPT 4o model but not the o1-preview or o1-mini models. Browsing has since become a standard part of ChatGPT, in ordinary chats and in the agents that work through tasks on their own.

ChatGPT Plugins (Replaced by Custom GPTs)

In 2023, plugins were another avenue to enhance ChatGPT’s capabilities. These add-ons could perform specialised tasks for a massive range of applications. For instance, a plugin could analyse your website’s SEO performance and suggest improvements, or another could assist in creating financial forecasts for your business.

Probably one of the most common applications in 2023 was web browsing, with plugins like BrowserPilot and WebPilot offering an alternative to Browse with Bing.

From November 2023, when OpenAI introduced custom GPTs, ChatGPT plugins were gradually phased out, and the plugins beta was wound down during 2024.

Custom GPTs let users easily create AI models tailored specifically to their needs, with a higher level of customisation and control than the publicly available plugins, which made them more versatile for both personal and business applications, and by the time OpenAI opened its GPT Store in January 2024 more than three million had been made, Tom’s Guide reported. The story has since come full circle. In 2026 OpenAI said it would retire custom GPTs in favour of plugins, a newer format built from skills and connected apps, with custom GPTs scheduled to stop running on 11 December 2026.

Advanced Data Analysis (formerly Code Interpreter)

The future of AI looked especially exciting in 2023 when we considered how powerful ChatGPT-4 was using Advanced Data Analysis (formerly Code Interpreter).

Imagine a plugin that could understand, write and evaluate programming code. That would be a major step forward for developers, data scientists, and analysts. Well, this is what Code Interpreter could do, and very effectively too. So much so that many developers were already using it in 2023 to speed up their coding workflows.

Powered by GPT-4, Code Interpreter could do so much more than coding, which is the main reason why it was rebranded to Advanced Data Analysis. With its wide range of Python libraries, ability to upload multiple files, and provide answers based on natural language prompts, there wasn’t much it couldn’t do.

For example, some popular use cases when we first wrote this in 2023 were:

1. Data extraction from PDFs: Read and analyse data from a PDF document.
2. Text parsing: Uses Python’s regular expressions (regex) to match and extract details, illustrating its ability to handle complex string manipulations.
3. Data aggregation: Correctly aggregating data from a range of different sources and even document types.
4. Real-time data analysis: Assist in real-time data analysis, automating repetitive tasks.
5. Educational support: Help in learning and teaching programming.
6. Debugging: Debug code and handle complex errors.
7. Mathematical calculations: Perform complex maths operations.
8. Data insights: Analyse raw business data and output reports and insights.
9. Charts: Create charts and graphs using structured and unstructured data.
10. Code testing: Run and test code in a sandbox, firewalled environment.
11. Natural Language Processing: Coding using natural language, allowing even non-programmers to use the power of code.
12. Code experimentation: A quick and easy way to experiment with code snippets.

If the above list wasn’t impressive enough, we’ve also written about how Screaming Frog’s SEO Spider tool and Advanced Data Analysis can be used to scrape entire websites into Excel form and even convert bulk Excel entries to individual Word documents using ChatGPT-4 to provide detailed recommendations and analysis.

OpenAI’s May 2024 post on improvements to data analysis in ChatGPT, published alongside GPT-4o, presented these features as part of ChatGPT itself rather than as a separate mode, with interactive tables and charts and uploads straight from Google Drive and Microsoft OneDrive, and that built-in data analysis is what replaced the Code Interpreter we used in 2023.

With GPT-4, the key to getting the most out of these features was prompt engineering. By the time GPT-5 arrived in August 2025, browsing, running code and analysing data had become things the model could do by itself as steps in a longer task, and it could reason through a complex request before acting on it, which made it feel more like a capable assistant than a sophisticated search tool. Neither GPT-4 nor any of its successors is AGI, but with the right approach, ChatGPT was, and remains, a powerful asset for any business.

AGI in 2024: OpenAI’s o1 and the Shift from Pattern Matching to Reasoning

We wrote the first part of this section in September 2024, and we’ve kept it as a record of how the o1 launch looked at the time. Following on from the adventures of Jimmy Apples and Q-star in 2023, things appeared to go relatively quiet in the first half of 2024.

Then in September 2024, OpenAI announced its new o1 models, which focus on enhanced reasoning for complex tasks in science, coding, and maths, with two versions: o1-preview (larger, broader knowledge) and o1-mini (faster, optimised for coding), available for ChatGPT Plus users.

The most intriguing parts of this announcement were the “more time thinking before they respond” and “reason through complex tasks and solve harder problems” comments.

There was no explicit mention of AGI but it sounded an awful lot like AGI and again raised the question: had AGI already arrived, or was it just around the corner? Looking back from 2026, o1 was a genuine step forward in reasoning, as the rest of this section explains, but it wasn’t AGI.

Is OpenAI’s ‘o1’ Model AGI?

The release of OpenAI’s “o1” model in September 2024 was a relatively quiet but very important one, and OpenAI’s decision to reset the naming scheme to “o1” signalled a fundamental shift in AI development. Unlike previous iterations like GPT-3 and GPT-4, which focused on scaling model size, the “o1” model represented a leap in reasoning and cognitive depth.

The reset marked a new chapter for OpenAI, with the emphasis moving from how large a model was to how well it could think, and the break from the GPT-x naming convention underlined the significance of these advancements, positioning “o1” as a groundbreaking step forward in AI’s evolution. OpenAI went back to GPT names with GPT-5 in 2025, but the shift in emphasis from size to reasoning stuck.

It marked a turning point that many in the industry had been waiting for, because previous models including GPT-3 and GPT-4 operated primarily through next-token prediction, a sophisticated form of pattern matching that produced impressively fluent text but, as we saw it in 2024, lacked genuine reasoning.

The question in 2024 was whether o1 itself was AGI or merely a stepping stone, and its advances made the case look stronger than for any model before it, starting with a new focus on Chain of Thought reasoning, where the system takes measurable time to “think” through complex problems step by step before responding, which is closer to how humans approach complex tasks. Rather than producing an instant response based on statistical probability, the model evaluated multiple paths, considered potential errors, and refined its output. This feature was a marked improvement over previous models like GPT-4, which were largely limited to single-pass responses without deep reasoning.

We saw the early signs of systems that can reflect on their answers, evaluate mistakes, and refine strategies much like a human would. Unlike traditional models that answer in a single pass, the “o1” model was trained with large-scale reinforcement learning to use its chain of thought, and OpenAI found its performance improves with more time spent thinking at inference time, which allowed the model to think and adapt as it solved problems and significantly improved its reasoning and output quality.

On advanced academic benchmarks, the “o1” model achieved PhD-level performance on GPQA, a set of physics, biology and chemistry problems, vastly outperforming its predecessors. According to OpenAI’s o1 post, it scored 83% on the 2024 AIME, a qualifying exam for the USA Mathematical Olympiad, using a consensus of 64 samples (74% on a single attempt, against GPT-4o’s 12%), and placed in the 89th percentile in Codeforces coding competitions. These results highlighted the staggering gap between GPT-4 and “o1”, particularly in tasks requiring deep reasoning and advanced knowledge.

While the “o1” model pushed AI closer to AGI, it still fell short of the full spectrum of human-like intelligence that defines AGI. That would have required the ability to generalise across all fields and perform a wide range of intellectual tasks autonomously, which the “o1” model, though impressive, couldn’t yet fully achieve.

With o1, OpenAI was laying the groundwork for its next generation of models, and the model proved to be a critical stepping stone: when GPT-5 launched in August 2025, chain-of-thought reasoning had become a core feature built directly into the system, and the reasoning approach of the o-series had effectively been absorbed into the main GPT line. Across the industry since then, reasoning has become something every frontier AI model is expected to do by default.

In 2024 we called this shift from pattern matching to genuine reasoning a major step toward AGI, and described it as the difference between a system that can produce convincing text and one that can actually think through problems; every subsequent model, including GPT-5.4 and GPT-6 Astra, builds on it. Looking back from October 2026, we’d qualify that a little. Reasoning models really do work through problems step by step, and on problems with checkable answers the results speak for themselves, but our own study found that they reason well inside a framework they’re given and still have to be told when a problem calls for a new one, which is why the reasoning test in our ten-test scorecard, further down, says only partly.

AI Developments: Our 2024 Predictions and What Happened

When we added o1 to this article in September 2024, we saw it as another inflection point for AI, with the conversation shifting away from scaling models like GPT-4 toward the deeper cognitive abilities that “o1” represented, and we used the moment to set down some predictions. At the time of writing them, the question was: what does the future of AI look like? Looking back, the answer has been even more dramatic than we anticipated, and the chart below shows how quickly the releases came.

Here are the original predictions for 2024-2025 as they were published, with status notes brought up to date in October 2026:

10 Positive Predictions

1. Widespread adoption of “o1” models: Expect to see the “o1-preview” and “o1-mini” models rolled out across more sectors, pushing the boundaries of reasoning in fields like coding, complex maths, and science. (Status, October 2026: o1 was widely adopted and then superseded. OpenAI followed it with the o3 line, starting with o3-mini in January 2025, and the reasoning approach was then built into GPT-5, which launched in August 2025.)
2. Custom GPTs become the new norm: Custom GPTs will dominate, allowing businesses and individuals to create AI models tailored specifically to their needs, increasing productivity and operational efficiency, while we will also see plugins for WordPress, AI content creators, and other external platforms using OpenAI’s APIs. (Status, October 2026: custom GPTs were widely adopted, with more than three million made by the time the GPT Store opened in January 2024, but, as described in the section above, OpenAI now plans to retire them in December 2026 in favour of plugins.)
3. Open-source AI takes centre stage: As projects like Meta’s LLaMA gain traction, the open-source AI movement will accelerate, giving smaller companies and individual developers more power to build sophisticated AI models. (Status: Meta released Llama 3 in April 2024, and it and other openly available models gained significant traction, narrowing the gap with proprietary models.)
4. AI becomes a key player in healthcare diagnostics: The enhanced reasoning capabilities of models like “o1” could be applied to diagnostics, drug discovery, and even personalised treatment plans. (Status: AI is increasingly used in diagnostics and drug discovery, with regulatory frameworks catching up across multiple countries.)
5. Augmented reality (AR) and AI merge: Expect further integration of AI with AR and VR, especially in consumer products like Meta’s new Ray-Ban smart glasses and Apple’s Vision Pro, offering more immersive, AI-powered experiences. (Status: Ray-Ban Meta glasses had sold about two million pairs by February 2025, according to EssilorLuxottica’s chief executive. Apple Vision Pro launched in February 2024 at $3,499 and sold far fewer, and in October 2024 Apple was reported to be cutting production.)
6. AI-driven automation in eCommerce: Look for more advanced AI solutions in inventory management, personalised shopping experiences, and logistics, as companies adopt AI to simplify and speed up operations. Businesses with professional eCommerce development will be best positioned to take advantage of these advancements. (Status: AI-driven personalisation and inventory management are now standard across major eCommerce platforms.)
7. The evolution of AI search engines: The rivalry between Google and Microsoft will heat up, with Bard and Bing competing to dominate AI-powered search and offering smarter, more accurate responses. (Status: Bard became Gemini in February 2024, and Bing Chat was renamed Microsoft Copilot in November 2023. AI-powered search is now a major battleground.)
8. Advanced AI in financial markets: AI models like “o1” could be used to navigate complex financial markets, offering predictive insights that go beyond traditional data analysis, leading to new fintech applications. (Status: AI-powered trading tools and financial analysis platforms have seen significant growth in adoption across institutional and retail markets. Prediction remains a separate problem from calculation.)
9. The rise of AI-driven decision-making: In businesses, AI will increasingly be relied upon for complex decision-making processes, shifting from task-based automation to strategic leadership roles. (Status: Agentic AI systems now handle multi-step business workflows autonomously, confirming this trend ahead of schedule.)
10. Potentially “o2” AGI or GPT-5 AGI: As models like “o1” push reasoning and learning capabilities forward, public anticipation for AGI will continue to build, possibly with new revelations or breakthroughs in 2025. (Status, October 2026: OpenAI released o3-mini, then GPT-5 and GPT-6 Astra, and no system has been verified as AGI; we look at the AGI claims made for Astra further down.)

4 Not So Positive Predictions

While AI has immense potential, there are concerns and challenges we may face in the near future:

1. AI-generated misinformation: As models like “o1” become more sophisticated, expect more AI-generated content to blur the lines between fact and fiction, making it harder to detect false information. (Status: AI-generated deepfakes and misinformation have become a growing concern, with detection tools struggling to keep pace.)
2. Educational challenges with AI: With students increasingly turning to AI tools like ChatGPT for homework and essay writing, traditional education methods will face disruption, leading to ethical and academic integrity concerns. (Status: Schools and universities continue to grapple with AI use policies. Some have embraced AI as a learning tool while others have imposed restrictions.)
3. AI’s impact on democracy: The 2024 election cycle was expected to witness new forms of AI-driven misinformation campaigns, sparking calls for stricter AI governance. (Status: AI-generated deepfakes and robocalls did feature in the 2024 US elections, and after a fake Biden robocall before the New Hampshire primary, the US regulator, the FCC, ruled AI-generated voices in robocalls illegal in February 2024.)
4. Greater AI regulation and compliance: As AI becomes more ingrained in society, governments will begin to impose stricter regulations around its use, especially in sectors like healthcare, finance, and cybersecurity. (Status: The EU AI Act came into force on 1 August 2024. Several countries have introduced or proposed AI-specific legislation, and some governments have gone further still, choosing to block access to advanced AI systems entirely over sovereignty and safety concerns.)

Most of these came true, several sooner than we expected, and the releases that followed, from GPT-5 in 2025 to GPT-6 Astra in 2026, showed that AI development was still speeding up.

GPT-5: Released, but Was It AGI?

We first wrote this section in March 2026, and it records how GPT-5’s arrival looked after years of build-up. GPT-5 was at the centre of enormous speculation throughout 2023 and 2024, with various reports claiming it would achieve AGI status, some of them, as we saw above, insisting it already existed inside OpenAI.

When GPT-5 finally launched on 7 August 2025, it was a significant leap forward but not AGI. Rather than a single monolithic model, GPT-5 launched as a dynamic system comprising a fast, high-throughput model and a deeper reasoning model, managed by a real-time router that decides which to use, as OpenAI’s GPT-5 launch post describes. It introduced integrated chain-of-thought reasoning, bringing the reasoning first seen in o1 into the main ChatGPT product, and agentic functionality (autonomously setting up desktops and browsing the web).

Measured against the definitions earlier in this article, GPT-5 was a real step up and still a long way short. On the Definition of AGI scale that Hendrycks, Bengio and colleagues published in October 2025, it more than doubled GPT-4’s score while remaining well short of a well-educated adult, and nothing about it came near OpenAI’s own charter bar of highly autonomous systems that outperform humans at most economically valuable work.

OpenAI kept iterating rapidly: GPT-5.2 (11 December 2025) and GPT-5.4 (5 March 2026) each added refinements to reasoning, coding, and computer control, and GPT-5.4 was the most advanced publicly available model when we updated this article in March 2026. By the summer the leading companies were giving their models names as well as numbers. GPT-5.6 Sol was OpenAI’s flagship by July 2026, when we set it against Anthropic’s premium model, Claude Fable 5, in our GPT-5.6 Sol vs Fable 5 comparison, and GPT-6 Astra overtook both in September.

So while GPT-5 showed that progress hadn’t stalled, it did not deliver AGI. The question then shifted to GPT-6, but before it arrived came a change in how AI was used, which the next section covers. Our piece on what we expected from GPT-5 before launch, linked below, records how the hopes for GPT-5 looked beforehand.

What is GPT-5 from OpenAI: When Is It Coming & What Can We Expect?

The Rise of Agentic AI: Systems That Take Action

GPT-5’s agentic features were part of what was perhaps the most significant development of 2025 and 2026, the rise of “agentic AI”: systems that plan and carry out complex tasks across different software environments as well as answering questions.

Traditional AI tools were reactive, in that you asked ChatGPT a question and it gave you a response, whereas systems built on agentic AI can be given a goal and then work autonomously to achieve it. They can browse the web, write and execute code, interact with APIs, manage files, and coordinate multi-step workflows without requiring human input at every stage.

When we added this section in March 2026, we wrote that it would change how businesses interact with AI, and by October that had started to happen. Instead of using AI as a faster search engine or writing assistant, some organisations were handing whole workflows to agents, from data analysis and report generation to customer service escalation and supply chain monitoring.

The companies building agents have been candid about their limits, too. In June 2025 Anthropic’s engineers published a write-up of a multi-agent research system, which says such set-ups use about 15 times more tokens than ordinary chats (tokens are the chunks of text a model reads and writes, and what its use is billed by), and that LLM agents are not yet great at coordinating and delegating to each other in real time, which is worth bearing in mind whenever a launch promises a team of agents that runs itself.

GPT-6 Astra: Has OpenAI achieved AGI?

With agents already doing real work, GPT-6 was the model many expected to settle the AGI question, and for a few days in September 2026 some of the people closest to it said it had. We’ve taken this part in order: what was expected before launch, what Astra showed once it arrived, and how the loudest AGI claims of the last three years hold up against their original sources.

What We Expected from GPT-6 Before Launch

We wrote most of this part in March 2026, before GPT-6 launched, and have kept it to show what people expected. By March 2026, with GPT-5.4 live and demonstrating increasingly sophisticated reasoning, the AI community’s attention had turned to GPT-6, anticipated for release later in the year.

In August 2025, shortly after GPT-5’s launch, OpenAI CEO Sam Altman had said that GPT-6 would come on a faster timeline than the two and a half years between GPT-4 and GPT-5, and that “the big breakthrough will be memory”, as The Decoder reported. OpenAI’s Stargate infrastructure project, announced in January 2025 as a $500 billion build-out of data centres and computing power in the US, was being built to train and run its next generation of models.

Expected advances in GPT-6 included deeper memory capabilities, more advanced agentic behaviour, and enhanced multimodal reasoning. These are precisely the capabilities that researchers identify as prerequisites for AGI, and whether GPT-6 would cross the AGI threshold or simply move the goalposts further was the defining question in AI development at the time.

What made this generation particularly significant was the convergence of several trends: chain-of-thought reasoning (introduced by o1 in 2024), agentic task execution (refined through GPT-5), and the massive scaling of compute infrastructure. If AGI is achievable within the current architectural paradigm, GPT-6 represented our best near-term opportunity to find out.

Is GPT AGI? What GPT-6 Astra Showed at Launch

GPT-6 arrived as GPT-6 Astra on 3 September 2026, so the gap between generations was indeed short, and the launch was pitched as if AGI had come with it: OpenAI’s president, Greg Brockman, told reporters he personally believed OpenAI had reached AGI and thought it might be about this model, closing the launch by welcoming people to the AGI era, according to Axios’s report that day. He hedged more than the headlines did, though, and Implicator’s report of the launch says he “described the change as a continuum rather than a single threshold”. Nothing published at launch claimed that the charter’s bar had been met, and OpenAI’s October maths results came from an unreleased internal model that outsiders can’t test, so on the evidence published so far our answer is that Astra isn’t AGI.

The independent measurements put its progress in a more modest light. In its benchmark of Astra, published on 9 September, Artificial Analysis, which runs the same battery of tests on every major model and combines them into one Intelligence Index score, gave Astra 53. The number means little on its own, since the index exists to compare models, but it put Astra six points ahead of its predecessor, GPT-5.6 Sol, and level with Anthropic’s premium model, Claude Fable 5.1. On the same organisation’s AA-Omniscience test of factual knowledge, Astra’s accuracy rose by four points and its hallucination rate fell from 92% to 51% at its maximum effort setting (effort controls how long a model spends reasoning before it answers, and the highest setting costs the most). That is a big improvement in one release, and it still leaves a hallucination rate on that test of about one in two. On ARC-AGI-3, the test of unfamiliar problems that OpenAI cited at launch, the result depended heavily on how the test was set up, as the claims below show.

Astra is the highest-scoring OpenAI model on that index, and it handles long agentic tasks better than its predecessors. ChatGPT, remember, is the product rather than the model: since GPT-5 it has routed each request to a fast model or a deeper reasoning one, so asking whether ChatGPT is AGI really means asking about the best model it can call on, which in October 2026 was Astra. Astra also took part in our own study, running inside OpenAI’s Codex for some of the idea generation, although we didn’t test it on its own. What we can’t find evidence of, in Astra or anywhere else, is the part of general intelligence that matters most in this article: a model learning from what happens to it, from one conversation to the next.

Has AGI Been Achieved? The Claims, Checked Against Their Sources

Claims that AGI had arrived peaked in September 2026, when GPT-6 Astra launched. Rather than repeat the headlines, we went back to what each underlying announcement, paper or post actually said, and took the six most widely shared claims in date order, oldest first.

“AGI has been achieved internally” (2023), false or unsupported: this is the story of Jimmy Apples, the anonymous X account covered in the 2023 section, where the phrase started on 18 September 2023 before turning up a week later in a Reddit comment attributed to Sam Altman, which was then edited to say it was a joke. Three years on, nothing matching it has been shown.

“AI already has human-level intelligence” (Chen and colleagues, February 2026), partly true: the Comment by Chen, Belkin, Bergen and Danks is a serious argument, and it holds under their definition, that LLMs already show general intelligence by the standards we apply to people; what it doesn’t show is learning from consequences or supplying new facts. (We read the summary and supplementary information, not the paywalled full text.)

“Welcome to the AGI era” (Greg Brockman, 3 September 2026), misleading: the quote is accurate, but no published transcript or OpenAI page attaches a test to it, and a personal belief stated at a launch isn’t evidence that a system transfers what it learns or sets its own goals.

“GPT-6 Astra scored 99.9% on ARC-AGI-3, so it’s AGI” (3 September 2026), misleading: this one is the most instructive of the six. ARC Prize runs ARC-AGI-3, a test built to measure how well a system copes with problems it hasn’t seen before, and its own post on GPT-6 Astra, published the same day, gives two sets of results. In its standard set-up, the one it runs the same way for every model, Astra’s best score was 62.7%, at maximum effort, for a reported cost of $26,098 across the test’s semi-private set of tasks. In a second set-up built on OpenAI’s own features, which lets the model keep its private working notes from one step to the next and condenses long conversations, it scored up to 99.9% at its “high” effort setting, and for less money ($18,817), presumably because carrying its earlier work forward left it less to redo. ARC Prize was explicit that it was not claiming Astra is AGI, prefacing that with a concession:

“…while we believe Astra represents meaningful progress towards generalization, we are not claiming that it is AGI.”

“AGI has arrived” (Jensen Huang, 6 September 2026), false or unsupported: Nvidia’s chief executive posted on X, as reported by Business Today on 7 September, that Astra was trained on Nvidia hardware, that AGI had arrived and that 400,000 more GPUs were coming online next, without citing any test. The company selling the chips announced the milestone in the same breath as its next batch of hardware, OpenAI led its own launch with the 99.9% score from a set-up built on its own features, and Astra’s price was 2.5 times that of its predecessor, GPT-5.6 Sol, according to Artificial Analysis. We don’t think that makes either claim dishonest, but it’s worth knowing that both companies stood to gain from the word that week.

“AI maths is improving 10x a month” (October 2026), misleading: the figure traces to a count given in an early-October YouTube video by the AI commentator Wes Roth, of about 10 published AI maths results in August, 100 in September and nearly 1,000 in early October. That is a count of how fast results were being published, which says nothing about how fast capability is improving, and OpenAI’s maths repository reports no peer-review status for them. It spread quickly all the same, and we repeated it ourselves before checking where it came from.

The table sums up the six claims. Each one was checked against its primary source on 10 October 2026.

Claim	Where it circulated	Primary source checked	Verdict
AGI achieved internally (2023)	The Algorithmic Bridge, 29 Sep 2023	Altman’s edited Reddit comment	False/unsupported
AI already has human-level intelligence	The blog 3 Quarks Daily, Feb 2026	Nature 650, 36–40	Partly true
Welcome to the AGI era	Axios, 3 Sep 2026	No published transcript; OpenAI Charter	Misleading
Astra 99.9% on ARC-AGI-3 = AGI	The AI news site Implicator, 3 Sep 2026	ARC Prize, 3 Sep 2026	Misleading
AGI has arrived	Business Today, 7 Sep 2026	Huang’s X post of 6 Sep 2026, as embedded	False/unsupported
AI maths 10x a month	Wes Roth’s YouTube video, about 8–9 Oct 2026	OpenAI maths repository	Misleading

Are today’s other frontier models AGI? (as of October 2026)

OpenAI isn’t the only company whose models get asked this question, so before turning to our own study we looked at the two main rivals, and at whether large language models as a family are on the path to AGI at all.

Is Claude AGI?

Anthropic, the company behind Claude, now sells more than one top model at once. Fable is the tier it described in mid-2026 as its most capable for demanding, long-running work, and Claude Fable 5.1 had followed Fable 5 by September. Opus is the other, and Anthropic’s announcement of Claude Opus 5.5, its newest model, released on 22 September 2026 as the first of a Claude 5.5 family, describes it as particularly good at long, sprawling jobs such as making the same change across a whole codebase or auditing one, without calling it AGI. Claude, mostly Opus 5.5 and Fable 5.1, did much of the best work in our study, so our verdict on Claude matches the one for ChatGPT: a capable contributor that still falls short of general intelligence.

Is Gemini AGI?

Google DeepMind’s own Levels of AGI framework, described near the start of this article, placed the Gemini of late 2023 at Level 1, Emerging AGI, with every higher level not yet reached, and we haven’t found a later placement for today’s Gemini models. We didn’t test a Gemini model in our study either, so we have no evidence of our own to add to the framework’s.

Are LLMs AGI, or the path to it?

Frank Emmert-Streib’s 2024 paper in Discover Artificial Intelligence, Is ChatGPT the way toward artificial general intelligence, argued that ChatGPT could not become AGI because it cannot act on its environment, has no policy of its own and lives only in text. Agents, computer use and image and voice input were built to remove exactly those limits, and the agents in our own study could act on a real computer and plan their own sub-tasks. So can LLMs become AGI? The 2024 objections were engineering problems and have largely been engineered around, whereas the gaps we measured, in truthfulness under pressure, learning from consequences and supplying new facts, haven’t closed at the same rate, and image input, which makes a model broader, does nothing for any of them. When we wrote up vibe coding vs agentic coding in June 2026, we found that every failed output led to a new written rule, which is a kind of learning, although it lived in our documents and the model itself never changed.

What we found when we tested it: 25 days, 276 ideas, no patentable invention

Agents and benchmarks only go so far, which is why we ran our own study in September 2026, the 25-day test that grew from a search for a power-assisted shopping trolley and ended up covering agentic AI as well as ordinary chat. It asked one practical question: can today’s best AI come up with something genuinely new and useful that nobody has done before? Whether AI can invent, and whether it can create new knowledge, each have their own article in this series, linked from the ten tests after this section, so here we concentrate on what the study found and what it means for AGI.

Why we ran the test

When I went looking for that trolley on 16 September 2026, for a family member who struggles with today’s trolleys, I found only powered walkers, priced between about £1,000 and £3,500, and the gap made me wonder whether AI could invent something genuinely new to fill it.

We soon realised that invention is a sharp test of general intelligence, because answering a question, writing code or solving a maths problem can all draw on what a model has already seen, whereas inventing something useful, and defending the claim that it’s new, means going beyond what is in the training data and on the web. A patent examiner asks exactly that: is this new, and would it be obvious to someone skilled in the field?

I came to it as a trained software engineer who tends to judge AI on the things I value in people (reading a situation, having an original thought and knowing when it’s wrong), and I also came in doubting that genuine creativity was a strength of LLMs, so the study was a chance to be proved wrong. As a company, Opace has used these tools daily since 2022, and by June 2026 we were building whole sites with AI agents, so if ChatGPT had quietly become AGI, an agency using it all day should have been among the first to notice.

How the test worked

The study ran for 25 days, from 16 September to 10 October 2026, and grew in stages:

• 16 to 26 September: ChatGPT, starting from the trolley search and running OpenAI’s GPT-5.6 in its Thinking mode.
• From 27 September: Claude, working through Claude Code (Anthropic’s tool that lets Claude read and change files on a computer), mostly on Claude Opus 5.5 and Claude Fable 5.1, alongside OpenAI’s Codex, mostly on GPT-6 Astra and GPT-6 Sol, with up to six conversations running at once. Our list of ideas reached its final 276 that same day.
• 27 and 28 September: a trial on one unrelated problem, to compare two models fairly (more on this below).
• 2 October: ten new ideas written down in full by OpenAI models working in Codex before any patent searching, so the search couldn’t quietly reshape them.
• 2 to 4 October: the agent team, mixing Claude and OpenAI models and working on business ideas rather than patents.
• 10 October: a debate with ChatGPT about the results, which closes this section.

The request itself set four conditions: an idea had to be genuinely new, patentable, cheap to test and commercially worthwhile. We then checked the ideas against existing patents and products, apart from 64 that were dismissed on sight without a formal search. To count as strong enough to justify a patent application, it had to pass five checks we could run at a desk, without building anything:

• New: nothing in the patents or products we could find already claimed it.
• Not obvious: we made five separate attempts, in the way a patent examiner would, to show it was an obvious step from what already exists, and it had to survive all five.
• Works in the worst case: even with every assumed figure 30% worse than expected, it still had to deliver at least twice the smallest benefit worth having.
• Can’t easily be designed around: if a rival could get 80% of the benefit another way for no more than 120% of the cost, the idea failed.
• Can make money: there had to be a credible route to a real return.

We set a firm rule early on that nothing would be built until we were confident an idea could be patented, which kept the study cheap and fast but also meant every judgement was made on paper (a limit we come back to below). The trial on one problem came about because shopping trolleys sit in a crowded field, and we wanted to know whether a different model would do better on a fresh problem, so we gave Claude Opus 5.5 and OpenAI’s GPT-6 Sol the same brief, finding the cause of damp in a home, and each produced nine ideas for the same checks.

What we found

The headline numbers, after merging duplicate ideas:

• Ideas: 276 distinct ideas, 274 of them generated by AI in about five hours on 27 September.
• Patents checked: 2,161 patent and application numbers.
• Strong enough to justify a patent application: none.
• Promising but unproven: 69.
• Ruled out: 207.
• Written down in full before searching: all ten were ruled out.
• The damp trial: none of its 18 ideas passed every check.

In plain terms, “promising but unproven” means an idea passed some checks but couldn’t be shown to pass them all without building or testing something, and “ruled out” means it clearly failed at least one.

What sits underneath those totals matters more than the totals themselves, because most ideas were new in their exact wording and very few failed simply because someone had already patented the same thing. They failed later, on whether they were inventive rather than an obvious combination, whether the physics would work in the worst case, how easily a rival could design around them, and whether they could make money. When the ten ideas written down before searching were checked across unrelated industries by underlying principle, rather than only within the obvious market, they met direct or near-direct prior art.

We also had Claude check our own write-up of the results for errors, and it found five statements that were wrong or misattributed, one of which had described the patent numbers we’d checked as patent families and so overstated how many separate inventions we had compared against. We corrected all five, and the conclusion stood.

How our agent team grew, and what it showed

Once the list of 276 invention ideas was complete, we tried agentic AI at Opace scale in early October as the last stage of the study before the debate with ChatGPT, and gave the agents a different job: to choose a business idea and start building it. We ran them in an AI agent workspace, software that lets several agents, each with its own role, share a project and message one another, and getting it running took longer than it should have, because Codex reported the set-up as complete when no agents were running (it accepted afterwards that it had been wrong). Claude then got it working by operating the computer directly, clicking and typing as a person would, and the faults turned out to be ordinary configuration problems, such as an expired login.

The team grew from 8 agents to 25 in less than a day, each given its own role, such as researcher or maker, and at one point Claude described the set-up as “finished and verified” shortly before eight more agents were added to it.

What the extra agents brought was more parallel searching and more claims of finished work for us to check. We asked whether 25 agents were too many for an agency of Opace’s size, and the agent we had set up to argue against the others’ proposals agreed that 25 was too many to run as the default team, although it advised keeping the specialist skills available for when they were needed. After two days the team had built one homepage, for a personal hobby project of mine, and Claude put the slowness down to “too much process for the output”, which fits the Anthropic finding, described in the section on agentic AI, that agents aren’t yet good at coordinating with each other. The judgement about what mattered still came from us.

What the models did well, and what they couldn’t supply

Reading the result as a simple failure of AI would miss what the models did well across all of these stages, which was to fill five roles: mapping the evidence, generating mechanisms, translating a need into patent-search terms, attacking weak claims and keeping an accurate record of what had been tried. They corrected our premises when we were wrong and, more often than we had expected given the complaints about sycophancy, reported results we didn’t want to hear, and ChatGPT itself wouldn’t promise a patentable invention on demand.

What none of them could do was produce a non-obvious idea when no new technical input went in. Over the 25 days we changed the method about 20 times, each time in response to the last failure: new prompts, new models, new rules and, at one point, an instruction to create ideas that didn’t exist in the models’ training data. That last one made the problem plain, because an instruction can’t make a model produce an invention nobody has ever documented. Every attempt drew on the same two sources, the models’ training and public web searches, so no new fact, measurement or experiment from outside the model ever went in.

When we listed the possible explanations for why no idea made it through, about 40 of them, the best supported were that the ideas weren’t inventive enough, that their physics didn’t hold up, and that nothing new went in. The least tested was that the models were simply trying to please us, which a blind test across three models could settle.

The strongest counter-evidence we know of comes from a blind study published at ICLR 2025, in which Si, Yang and Hashimoto found that expert reviewers rated LLM-generated research ideas as more novel than ideas from human experts, by a margin unlikely to be chance, though slightly weaker on feasibility. Ideas that read as novel are a different thing from inventions that survive a patent search, but the two findings sit side by side, and the fairest summary of our study is that AI proved a valuable contributor to invention while remaining unproven as a sole inventor.

What the test can’t tell us

Our study has real limits, and the biggest is that it was a single company’s study, run at a desk, with no lab and no private data. AI did the patent searching, so a model’s patent checks stood in for a professional search and are not legal opinions, there was no physical test, and the samples were small. The first areas we searched, including mobility aids, consumer health and home diagnosis, were already crowded with patents, and a gap in the market isn’t the same as a gap in the patents. The checks also varied in depth: 64 ideas were dismissed on sight without a search, the damp trial covered one problem on two models, and the ten ideas written down first all came from OpenAI models working through Codex.

Small samples also limit what those zeros can tell anyone: with no successes in ten tries, the true success rate could still be as high as about 30%; with none in 18, it could be up to about 17%; and even the 276 ideas only rule out a rate above about 1.1%, a figure that is optimistic because so many of the ideas are variations on each other. What the study shows, then, is that this approach didn’t produce a patentable invention, and it doesn’t prove AI can’t invent.

Our own method had weaknesses as well, starting with the rule against building anything, which meant only advantages you could calculate at a desk could pass, which is a choice we made rather than a limit of LLMs, and the damp trial’s brief accidentally showed the models some of our existing ideas before they wrote their own. The study used chat models and AI agents, not the purpose-built systems with automated checkers behind 2026’s maths results, so it says nothing against those, and although GPT-6 Astra did some of the work inside Codex, we didn’t test it on its own, or test Gemini at all.

A fairer test would fix the method in advance, use several problems whose physics can be calculated, use at least three different families of models, and check whether the process can rediscover recently granted patents from the need alone. It would also add the input ours never had: private data, an expert’s records of what has failed before, a measured anomaly or a real experiment.

Opace sells AI consulting, so we aren’t neutral, and two of my personal hobby sites, which are separate from Opace, bear on the argument as well. After the study I built PennyForYours, a hobby project for getting quick reactions from real people, which only makes sense if AI still needs people. I also run OutsmartMoney.ai as a hobby site, using AI to help explain markets without claiming to know where they’re going (it’s educational and not financial advice, and its own pages warn that AI summaries can be wrong). We’ve tried to let the evidence make the argument, and we mention them so you can weigh it with that in mind.

The 46-minute debate where ChatGPT changed its mind

Our experiment into whether AI can invent came to its conclusion on 10 October 2026, when, prompted by Wes Roth’s YouTube video about OpenAI’s new maths results (the source of the “10x a month” figure checked earlier, which I repeated in the conversation before checking it), I put the AGI question to ChatGPT directly in a conversation that lasted 46 minutes. We didn’t record which model ChatGPT chose to answer with. My central point was the one this whole series keeps returning to:

“Reasoning is great, but you can only reason based on what you know.”

Asked whether an LLM could produce genuinely new ideas, ChatGPT opened with a firm yes: “Yes. An LLM can produce genuinely new ideas, solutions and discoveries that did not exist in its training data, using reasoning alone.” It backed that with real evidence, from Google DeepMind’s FunSearch to OpenAI’s October maths release, and when it later summed up where that left AI, it was more careful, rating the generation of new ideas as “demonstrated” but a general intelligence that keeps improving itself through its own feedback loop as “not yet established”.

I then described our study: hundreds of ideas, thousands of patents checked and none we could confidently call unique. Over the rest of the conversation ChatGPT moved towards my position four times, accepting that our result challenged its claims and calling my central argument logically sound. The clearest slip involved a false memory of mine. I said I had asked it to explore new physics earlier in the project, and within seconds it agreed, accepting in the same reply two other points I’d made (that existing patents don’t prove a physical limit, and that an intelligent system shouldn’t need to be told how to think): “You’re right on all three points.” I can find no record that I did; the closest words in my chat history are about scientific thinking, physics and conservation of energy.

It held its ground once, after I said it seemed to be telling me what I wanted to hear. ChatGPT admitted it had been adjusting its explanations towards each new objection, then disagreed with me for the first time: “extraordinary mathematical performance does not establish an equally extraordinary capacity for open-ended invention.” It also proposed a fairer test, one we’d like to see run: the same invention brief given blind to three models built in different ways, with the results judged independently.

For the AGI question, the debate showed something the benchmarks miss: ChatGPT had the knowledge and reasoning to argue either side convincingly, but no view of its own, anchored to evidence, that held when I pushed, and at the moment it mattered most it agreed with my false memory when a general intelligence, we think, would have checked.

The ten-test scorecard

Our study tested one ability in depth, but general intelligence isn’t one ability, so we split it into ten questions, each of which has its own article in this series. Here we give our answer to each, with outside evidence and, where we have it, something from our own study.

Test 1, does it tell the truth: not yet, and OpenAI itself bore out the sycophancy research described earlier when it rolled back an April 2025 GPT-4o update for being too sycophantic. Our study saw both sides, since the models reported unfavourable results more often than their reputation suggests, but ChatGPT also agreed with a false memory of mine without checking, and AI tools twice reported set-up work as complete before it was. Read more on whether AI lies or just gets things wrong.

Test 2, can it do maths: this is the one we mark as a pass, on checkable problems. The public evidence was strong long before 2026, with o1’s competition-maths results in September 2024 and every frontier model since building on them, and the newest evidence is more striking but harder to judge. OpenAI published about 720 maths write-ups on 6 October 2026, covering 372 distinct results, from an unreleased internal model that outsiders can’t yet test; they are at different stages of checking, and roughly 42% of the main results have been verified in Lean, software that checks a proof step by step. Early ChatGPT couldn’t reliably do simple calculations, which led many people to conclude that LLMs weren’t built for maths at all, but the more useful split turns out to be between calculating, which models now do very well, and predicting, such as a future share or Bitcoin price, which nobody can do reliably. More in our piece asking can AI do maths at research level.

Test 3, can it reason: partial, because although we called OpenAI’s o1 an AI inflection point when it arrived in September 2024, and reasoning is now built into every frontier model, the gap is in knowing when and how to apply it. In our study the models reasoned soundly inside the problem as it was framed, and the ideas they produced were, almost without exception, new combinations of familiar mechanisms; none of them changed the framing itself. That suggests today’s models are much better at solving hard problems within an established framework than at finding a new one. See whether AI can reason or only mimic reasoning.

Test 4, can it think for itself: not yet, because although agents can now plan and carry out long tasks, every goal in our study came from a person, along with the rules and the decision to stop, and although the agents broke goals into sub-tasks well, none of them proposed a goal of its own. More in whether AI can think for itself or only follows instructions.

Test 5, can it learn: partial, because it learns within a conversation but not, on its own, from experience. Claude can save a standing instruction to its memory, and in our study it saved one asking it to push back and stay impartial, yet days later it still gave up its own research-based advice on how big an agent team should be as soon as we disagreed, so the rule had been stored and the behaviour it asked for didn’t follow. Hendrycks and colleagues found long-term memory the weakest part of today’s systems, and in our study the memory that mattered lived in files we wrote around the models. The detail is in whether AI can learn from its own experience.

Test 6, can it invent: partial, meaning a contributor yes and a sole inventor unproven. Our 276 ideas produced nothing strong enough to justify a patent application, and we’ve still marked it partial rather than not yet, because the models did real inventive work. They generated, searched and filtered ideas far faster than a person could, and in Si, Yang and Hashimoto’s blind study, expert reviewers rated LLM research ideas as more novel than human experts’ ideas; what none of them did was produce an invention on its own. The law hasn’t moved either: the UK Supreme Court held in Thaler (2023) that an inventor must be a natural person. See whether AI can invent something patentable.

Test 7, can it create new knowledge: partial, and to see why it helps to split “new” into three kinds (ChatGPT suggested the split in the debate that closed our study, and it’s a useful one): recombination, which puts existing ideas together in a new way; reasoned discovery, which works out something that follows from what’s known, as a proof does; and empirical discovery, which establishes a new fact about the world. LLMs do the first two well. The third needs evidence from the world, which was the input our study never had, and our 276 ideas were, in the main, recombination. Read can AI create new knowledge or only recombine it.

Test 8, how smart is it overall: partial, and jagged, which is the word Hendrycks and colleagues’ scores suggest too. We saw the same unevenness at a smaller scale. Claude recovered a complete copy of a long ChatGPT conversation in about five minutes after another tool had failed to save it, and the same assistant later described a set-up as finished and verified when it wasn’t. That unevenness is why we argued in our July 2026 comparison of GPT-5.6 Sol and Claude Fable 5 that businesses “buy outcomes, not benchmark prestige”. More on how smart AI really is once you look past benchmarks.

Test 9, is it smarter than humans: at many tasks, yes, and the most striking evidence is conversational. In Cameron Jones and Benjamin Bergen’s 2025 Turing test, GPT-4.5 with a humanlike persona was judged to be the human 73% of the time in five-minute chats.

What it doesn’t yet match is a person’s judgement of context. In our study an OpenAI model (GPT-6 Astra) rated a draft homepage built by our agent team as “fairly premium”, while the first person who saw it cold couldn’t tell what the business did and guessed “it’s something to do with parcels being delivered”; the model had judged how the page looked, and the person whether it made sense. Our agents suggested that the remaining limits may have less to do with cleverness than with what AI lacks outright (a body, a legal identity, money and liability), which seems to us worth testing, and so we mark this one partial. More on is AI smarter than humans at the things that matter.

Test 10, will it cure cancer: not yet, and not on its own, although this is where purpose-built systems show most clearly what narrow AI can do. In 2020 Demis Hassabis and John Jumper presented AlphaFold2, which has since predicted the structure of virtually all 200 million proteins researchers have identified and been used by more than two million people in 190 countries, according to the 2024 Nobel Prize in Chemistry announcement, which gave them half the prize for protein structure prediction. The prize recognised a purpose-built protein model, and a cure still needs laboratory work, clinical trials and regulators’ approval, none of which a chat model can supply on its own. The evidence is in our article on AI’s realistic role in curing cancer.

That makes one Pass, six Partials and three Not yets. We don’t mean any of the Partials as a polite no, since each marks something the systems genuinely do in some conditions that a general intelligence would do in all of them.

How close is AI to AGI?

With the study and the scorecard in mind, the obvious next question is how far away AGI might be. Nobody can give a date that doesn’t depend on the definition, but progress on real tasks can be measured, and the best measure we know of is METR’s.

METR, the independent research group introduced earlier, tracks the length of tasks that AI agents can complete with 50% reliability, measured by how long the same tasks take skilled people. In its March 2025 report it said that this “has been doubling approximately every 7 months for the last 6 years”. Its January 2026 update, Time Horizon 1.1, put the doubling time at about 196.5 days across all models, 131 days for models since 2023 and 89 days since 2024, while warning that the trend is somewhat sensitive to which tasks are included and that its confidence intervals are wide. A doubling time of 89 days means the length of task an agent can finish roughly doubles every quarter, so on that measure progress is fast and speeding up.

Our own view, from the 25-day study rather than a benchmark, is that the progress is real and uneven. The gaps in truthfulness, learning and invention aren’t closing at the speed of the METR curve, because longer tasks don’t make a model more truthful under pressure, better at remembering last week or able to supply a fact it was never given. So how far are we from AGI? On task length, closer every few months, but on learning from consequences and supplying new facts we haven’t seen a comparable measured trend, and without one we don’t think anyone can give an honest date.

Updated AGI Timelines: What Experts Predict in 2026

The timeline for achieving Artificial General Intelligence seemed to shrink dramatically between 2023 and early 2026, as the rapid scaling of compute power and the breakthrough of chain-of-thought reasoning models raised expectations, but the published expert forecasts are further out than the ones we quoted in March 2026. According to Our World in Data’s summary of AI timelines, half of the 356 experts who gave complete answers in Grace and colleagues’ 2022 survey put a 50% chance of human-level AI before 2061, the average expert in Zhang and colleagues’ survey of machine-learning researchers put it at 2060, and the Metaculus community forecast for the date of AGI gave a 50/50 chance by 2040 in November 2022.

Our March 2026 update said instead, citing the same Our World in Data page, that fifty percent of leading AI researchers predicted early forms of AGI between 2027 and 2033 and that prediction markets put weak AGI at 2029, and when we went back to the page it didn’t say that, which is why we’ve corrected it. Those forecasts were all made before reasoning models arrived. What has clearly sped up since is measurable progress on tasks, as METR’s figures above show, and much of that comes from the shift from models that simply predict the next word to systems that reason through several steps.

We also wrote in March that the compute used to train frontier models doubles roughly every six months, and that one holds up: Epoch AI found the training compute of frontier models grew about 4–5 times a year up to May 2024, which works out at a doubling roughly every five to six months. The industry is investing at a scale that suggests genuine confidence in near-term breakthroughs, although we’d read that as a measure of what investors expect, which isn’t the same as evidence of when AGI will arrive.

GPT-7 and Beyond

After Astra, the natural question is what OpenAI does next, and on GPT-7 there is very little reliable to report. When we checked while preparing this update in October 2026, we found nothing from OpenAI that names, previews or dates a GPT-7. A Pluralsight guide to GPT-7, published on 10 September 2026, reached the same conclusion; it rated a claimed August 2026 leak about a new pretraining run as unverified, because it rests on a single anonymous source, and estimated from the gaps between OpenAI’s past releases that GPT-7 might arrive in the second half of 2027. That estimate is an extrapolation from OpenAI’s release pattern, and OpenAI itself has said nothing that would confirm or rule it out.

What we can say is what GPT-7, or any rival model, would need to show for this article’s answer to change. On the evidence above, it would have to learn from consequences between one conversation and the next, hold a justified position when a user pushes back with nothing new, and bring in a fact about the world that nobody had given it, which are the three gaps in our verdict below. A higher benchmark score on its own wouldn’t settle it, for the reasons the ARC-AGI-3 results showed.

Artificial Intelligence Beyond 2026

AI has evolved at a blistering pace over the past three years, but where do we go from here? The future of AI could be defined by deeper integration into our daily lives, ethical challenges, and the growing influence of AGI.

As we move through 2026 and beyond, the following areas will likely dominate the AI landscape. Treat them as possibilities rather than forecasts.

AGI and Eventual Human Augmentation

When we first wrote this section in 2023, we said that, as surreal as it sounds, we weren’t lightyears away from human-AI augmentation, and that with the likes of Neuralink, Elon Musk’s brain-interface project, combined with breakthroughs in the AGI space, it wouldn’t be long until we could augment and enhance our cognitive abilities via artificial systems.

Neuralink was then focused on trying to treat patients that have issues with communication by placing an implant inside, or on, the brain. Early tests looked promising, and in May 2023 the company announced US approval for its first human trial, aimed at people with severe paralysis. The first participant received an implant in January 2024, according to Neuralink’s own progress update, and has since used it to control a computer.

It is easy to imagine a time when Neuralink can be inserted into the skull in order to enhance human capability rather than being used to fix cognitive or communication problems, but three years on we’d put “it won’t be long” less confidently, because the scenario assumes the AI half is general, and on the evidence in this article it isn’t yet.

More Capable Robots and AI Hardware

By 2023 we had seen the Japanese robots that existed in some hotels (although they were still controlled by humans at the time), and hotels such as Henn na Hotel Asakusabashi were appearing, where robots and AI play a key role in running the business and serving guests. When we checked in October 2026, the hotel’s own site still said that robots work at its front desk and throughout the building.

In 2023 we called this a gimmick with some utility and functionality, and expected to see a more comprehensive implementation of robots in the coming years, with robots able to operate with complete autonomy. A body matters to the AGI question, though it’s only one of the things today’s AI lacks, as the scorecard’s test of whether AI is smarter than humans suggests.

ASI: What About Artificial Super Intelligence?

As we noted when defining it alongside AGI near the start of this article, ASI is still purely hypothetical and highly speculative as of October 2026. However, with the rate of innovation and change taking place within the field of AI, we can see a future in which ASI becomes possible, although nothing yet shows it is close and we can’t find good evidence for putting a date on it. The nearer question, whether today’s AI is already smarter than humans at the things that matter, has a more useful answer, which Test 9 of our scorecard gives. The closest thing to ASI-like behaviour today is narrow: OpenAI’s October 2026 maths results may well be superhuman in places, but superhuman ability in one checkable field is still, by definition, narrow.

Self-improvement is the step that would change that picture, and research on it exists, though at an early and narrow stage. One example is the Darwin Gödel Machine (Zhang, Hu, Lu, Lange and Clune, 2025), a coding agent that rewrites its own code and keeps the versions that score better, raising its score on SWE-bench, a test of fixing real software bugs, from 20.0% to 50.0%, and on Polyglot, a coding test across several programming languages, from 14.2% to 30.7%. The underlying model stays frozen throughout, so the gains belong to the coding agent built around it, and the model’s general intelligence doesn’t change at all.

How Businesses Should Prepare for AGI Today

Whatever the dates turn out to be, agentic AI systems are already in practical use, so the real question for businesses is how to prepare for more capable AI, whether or not it ever earns the AGI label.

Here are the practical steps that forward-thinking organisations should be taking in 2026:

1. Audit your current AI readiness: Review your existing technology stack, data infrastructure, and team capabilities. Identify where AI is already being used and where gaps exist.
2. Invest in clean, structured data: AI systems, today’s and whatever comes next, will depend on high-quality, well-organised data. Businesses with messy or siloed data will struggle to take advantage of advanced AI capabilities when they arrive.
3. Build flexible architectures: Monolithic systems that cannot easily integrate new AI services will become a liability. Invest in modular, API-first architectures that can adapt as AI capabilities evolve.
4. Upskill your team: Prompt engineering, AI workflow design, and understanding of AI safety are becoming essential skills. Invest in training now rather than scrambling to catch up later.
5. Establish AI governance frameworks: Whether or not AGI arrives, organisations need clear policies around data privacy, algorithmic bias, and decision-making transparency. Getting these right early will build trust with customers and regulators alike.
6. Start with practical use cases: Rather than waiting for AGI, begin implementing agentic AI in specific workflows today. Each successful deployment builds organisational capability and confidence.

From our own 25-day study we’d add one more step, which is to write down what finished means before an agent starts and to check it yourself, because two of the clearest failures we saw were claims that set-up work was complete before it was.

Integrating these advanced models into existing platforms requires strong technical foundations, which custom web development solutions can provide to ensure your systems are ready for the future of AI-powered automation. If your website is where you’d rather start, that’s what our AI web design services are for. As AI consulting experts in Birmingham, we help organisations nationally and internationally prepare their infrastructure for the next generation of reasoning models, and in our experience a business doesn’t need to settle the AGI label before using today’s systems. What it does need is a bounded task, a check that catches the errors that matter and a named person who can stop the work, which is where our AI strategy consulting starts, before any choice of model.

Search is one place where the change has already arrived. If AGI does arrive, we’ve looked separately at what AGI would mean for SEO and search, but you don’t have to wait for it, because AI is already changing SEO. ChatGPT and Gemini already answer questions about businesses like yours. Checking and improving those answers is what our AI SEO agency work covers, alongside the SEO services that still decide where you rank in ordinary search results.

The last lesson we took from the study is about people, and it comes from the homepage our agent team built, where the first person to see the page cold caught, within 30 seconds, a problem a model had missed. If what you’re building depends on how people react, it pays to ask people, and that is the idea behind PennyForYours, the hobby project I built after the study for getting quick, honest first reactions from real people. It’s open by invitation while it’s tested with a small group.

The Future of AI: Final Thoughts and Our Verdict

The future of artificial intelligence holds extraordinary promise, but it also demands thoughtful preparation, and part of that is being clear about terms: ANI, AGI and ASI often cause confusion, and distinguishing between these stages of AI development is more important now than ever.

The shift from pattern matching to chain-of-thought reasoning, which began with OpenAI’s o1 model in 2024, turned much of the old speculation into something businesses could use, and agentic AI systems that can plan, reason, and execute multi-step tasks are already changing how businesses operate.

While models like ChatGPT and GPT-6 Astra are powerful, they are still limited in scope compared to true AGI. Whether the breakthrough comes from OpenAI, Google DeepMind, or another organisation entirely, we are closer to witnessing systems that can act autonomously across a broad range of tasks than at any point in history, and still some way from systems that can learn from what happens to them.

The debate about whether AGI already exists, fuelled by figures like Jimmy Apples, the anonymous X account behind the 2023 rumours, captivated the world for years. We initially predicted in 2023 that AGI was probably already developed behind the scenes at OpenAI, even though ChatGPT itself wasn’t AGI. Instead of a GPT-5 AGI release, we got the “o1” model, which we said at the time sounded an awful lot like AGI, and the subsequent focus on agentic AI systems has been even more transformative for businesses. That 2023 call turned out to be wrong, and it was three years of daily use, and a 25-day study of our own, that changed our minds.

For us to achieve “true AI”, we need AGI. It’s fascinating that with reasoning models and agentic AI tools, it can feel as if we’ve nearly arrived, even though the evidence says we haven’t. In 2023 we wrote that true AI or AGI can only be defined as a system that operates autonomously across a broad range of tasks, and that this seemingly minor detail meant we weren’t quite there yet. Agents have since come a long way towards that kind of autonomy, so it no longer settles the question on its own, and what keeps today’s AI short of AGI is the set of gaps in our verdict below. Ultimately, the future of AI will depend on how we prepare for the changes, benefits, and risks introduced by companies like OpenAI.

So, is ChatGPT artificial general intelligence? As of October 2026, our answer is no. Of the five positions set out near the start of this article, only two say yes in any sense. Chen and colleagues’ Comment says yes on the standards we apply to people, and Google DeepMind’s levels place today’s chatbots at “Emerging AGI”, the lowest level, which the framework equates with an unskilled person; we read that as a label for breadth, not for general intelligence in the sense most people mean. OpenAI’s own charter bar of doing most economically valuable work better than people isn’t met, the Definition of AGI scores leave GPT-5 well short of a well-educated adult, and ARC Prize, whose test OpenAI cited, declined to call Astra AGI.

ChatGPT is a capable contributor, broad, fast and at or beyond expert level on many checkable problems, but it isn’t a general intelligence, because it doesn’t hold a justified position under pressure, learn from consequences between one conversation and the next, or supply the new fact that invention needed. We’d change our view on the evidence of a public, reproducible system that held its position against a user with new evidence, carried a lesson from one real outcome into an unfamiliar task, or produced an invention that cleared independent checks. It’s a narrow verdict, built on one study and the published evidence, and we’ll update this article if the evidence moves.

What do you think about the future of AI? Will AGI arrive with the next model, or will it take years before we see its full potential? Please let us know in the comments section below or on social media.

Frequently asked questions

Is ChatGPT artificial general intelligence or narrow AI?

Neither label fits cleanly, because ChatGPT isn’t narrow in the way a chess engine or a spam filter is narrow: the same model drafts contracts, writes code, explains tax rules and plans a holiday. But it falls well short of OpenAI’s charter definition of AGI, systems that outperform humans at most economically valuable work. If you need a single word for it, broad is more accurate than either narrow or general.

How uneven that breadth is shows up in the independent measurements described in the GPT-6 Astra section. The model that tied for first on Artificial Analysis’s index in September 2026 still had a hallucination rate of about one in two on the same organisation’s test of factual knowledge, so a system can sit at the top of a league table and be unreliable on facts at the same time, which is, to us, what broad but not general looks like in practice. People sometimes ask whether adding image and voice input has made it general, and it has certainly made ChatGPT broader, since it can now read a chart or answer a spoken question, but it does nothing for the gaps that keep it short of general intelligence, which are about reliability, learning and new facts rather than about how many kinds of input it accepts.

Our own study adds a reason this matters most for new ideas, because one weakness we found in our process, looking back, was that an idea’s absence from a model’s memory had been treated as a possible sign that the idea was new. A model that treats the edge of its own knowledge as the edge of what exists will mistake a gap in its training for a gap in the world. Failing to find an excellent invention didn’t prove none existed either, only that our method hadn’t found one. So we use ChatGPT as a broad, fast collaborator, and check anything it calls new against patents, products and published work, whatever confidence it expresses.

Will ChatGPT become AGI?

That depends on which AGI you mean. On DeepMind’s levels, a later ChatGPT could move from Emerging to Competent by matching the median skilled adult on most tasks, and METR’s measurements suggest the length of task agents can finish is doubling every few months. On a definition that requires learning from consequences and supplying genuinely new knowledge, the path is less clear, because what would have to change is less about size and more about memory that updates from real outcomes, a stable view under pressure and access to new evidence from the world. GPT-6 Astra, released in September 2026, improved on agentic work and still didn’t show the first two, which suggests a bigger model trained on more text isn’t enough on its own. We’d watch for a public, reproducible result where a model carries a lesson from one real outcome into an unfamiliar task, without a person rewriting its instructions in between.

When will AGI arrive?

Nobody can give a reliable date, and the forecasts set out in the timelines section are further out than the headlines suggest, with surveys of experts in 2022 putting an even chance of human-level AI somewhere between 2040 and 2061 (we used to quote 2027 to 2033, which the source we cited doesn’t support). What has sped up since is measurable progress on tasks: for models released since 2024, METR found the length of task an AI agent can finish doubling roughly every three months, although it warns that the trend is sensitive to which tasks are measured. So the honest answer to “when?” depends on the definition. A system that can handle most office tasks for hours at a time may be close; one that learns from the world as a person does, and brings back facts nobody has recorded, has no date anyone can defend. If you’re planning a business around it, we’d plan for steadily more capable agents arriving every few months rather than for a single moment when AGI switches on.

Is AGI the same as superintelligence?

Not quite, although the two are often confused. AGI usually means human-level ability across most cognitive tasks, roughly as capable as a skilled person in general, while artificial superintelligence (ASI) means far exceeding the best humans in almost every field. Today’s systems are superhuman in a few narrow, checkable areas such as competition maths, which is impressive but still narrow, and since no current system improves its own general intelligence, ASI remains hypothetical.

Has AGI been achieved yet?

Not on the evidence publicly available when we finished this update, and as the GPT-6 Astra section shows, even the people closest to the strongest claims hedged more than the headlines did. OpenAI’s president, Greg Brockman, described a continuum, and ARC Prize, whose test OpenAI cited, said plainly that it wasn’t claiming AGI while acknowledging “meaningful progress towards generalization”. Both describe gradual progress towards something.

The independent numbers point the same way. Astra came out level with a rival on Artificial Analysis’s index, which isn’t what the arrival of a new kind of system would look like, and Google DeepMind’s levels make the gap concrete, because the step above Emerging AGI is Competent AGI, performance at least at the 50th percentile of skilled adults across a wide range of non-physical tasks, including learning new skills, and the framework listed that level as not yet achieved. We haven’t seen anything published in the month since Astra’s launch that shows a model clearing it across the board.

Our own study gave us a smaller, more specific reason for caution. The ten ideas we wrote down before any searching all looked new at first, and it was only when the searches went beyond the obvious market, into unrelated industries that used the same underlying principle, that each one turned out to have been done before, at least in part. A system can look as if it has reached something new simply because nobody has checked far enough, or in the right places, so the habit we’ve taken from the study is to ask who checked an AGI claim, how far and with what, before asking how impressive the headline number is. If a lab does get there, we’d expect the evidence to look less like a statement at a launch and more like the tests in our scorecard being passed in public, by a system that outsiders can run and check for themselves.

Does passing an AGI benchmark prove general intelligence?

A benchmark score can’t prove general intelligence on its own, and the people who build the benchmarks are usually the first to say so, because a benchmark measures performance on its own tasks under its own rules, and small changes to those rules can move the score a long way. ARC Prize’s two results for GPT-6 Astra, described above, show it clearly: the same model went from 62.7% to 99.9% when the set-up let it keep its private working notes between steps, and ARC Prize still declined to call it AGI. A benchmark can show that a system has a capability, although it can’t show that the system will apply that capability unprompted, remember what it learned or stay honest when a user pushes back, which is where our own study found the biggest gaps. The habit we now follow is to ask what set-up produced a number before asking what the number means, and to check whether the people who built the test agree with the headline.

What did your 25-day test actually show?

It showed that ChatGPT, Claude, Codex and, at the end, an agent team of as many as 25 could generate and screen ideas at remarkable speed, and that none of 276 ideas was strong enough to justify a patent application. Most were new in their exact wording and failed instead on whether they were inventive, whether the physics worked, how easily they could be designed around or whether they could make money. The missing input was a new fact or observation from outside the model. One study of one approach can’t prove that AI can’t invent, but it does show where this approach fell short, and what a fairer test would need to add.

Can you just ask ChatGPT whether it is AGI?

You can ask, and the answer can be interesting, but it isn’t evidence. In the debate with ChatGPT that closed our study in October 2026, the same ChatGPT that judged a self-improving general intelligence “not yet established” also shifted its position towards mine four times, so we’d judge it by tests, including tests of whether it holds its view under pressure, and give its description of itself very little weight.

Could ChatGPT be AGI for some people and not others?

In practice, it can feel that way, because its abilities are jagged. Someone who uses it for drafting, coding or exam-style maths meets a system that is often better than the people around them, and calling that general intelligence is understandable. Someone who needs it to hold a position under pressure, remember a lesson from one conversation to the next or find something nobody has written down meets its weakest side. The model is the same in both cases, so we think the fairer answer depends on the task you’re asking it to do.