No, artificial general intelligence has not arrived by any widely accepted definition. OpenAI’s president said in September that it is reasonable to feel the industry is in an “AGI era”, but Google DeepMind’s chief executive, several academics and the ARC Prize Foundation dispute or decline to make that claim. Much of the disagreement comes down to definitions.
Key takeaways
- “AGI” has no agreed definition, which is the main reason people disagree about whether it has arrived.
- Sam Altman has said, according to a TIME report summarised by The Decoder, that he expects an internal system he would call AGI by the end of 2026. Demis Hassabis has said today’s systems are nowhere near human-level AGI. Both are opinions or forecasts, not results.
- Independent experts quoted since the GPT-6 Astra launch say that high benchmark scores alone do not establish general intelligence.
- The International AI Safety Report (February 2026) found no significant overall effect on employment so far, but flagged possible effects on junior workers in exposed jobs.
- For everyday users, the practical advice is the same whoever is right: use the tools, check the output and keep your own skills current.
What do AGI and superintelligence actually mean?
AGI (artificial general intelligence) is usually taken to mean AI that can do most intellectual tasks at least as well as a person. Superintelligence means AI well beyond human ability across the board. Neither has a precise, accepted test.
OpenAI’s charter, as quoted by The Decoder, describes AGI as highly autonomous systems that outperform humans at most economically valuable work. That is a business-oriented definition. Others prefer cognitive definitions, such as matching human adaptability on unfamiliar problems. When two people use different definitions, they can look at the same system and honestly disagree.
What has actually happened?
- On 3 September 2026 OpenAI unveiled GPT-6 Astra. OpenAI says it meets the “Critical” threshold for cyber capabilities under its own Preparedness Framework, according to Fox Business.
- Access was staged: first a limited group of organisations, including members of a cyber-focused programme, then paying ChatGPT users, the API and Amazon Web Services (Fox Business, 9to5Mac).
- OpenAI describes Astra as state of the art in areas such as computer use, software engineering and professional work. Those benchmark results are the company’s own.
What do the labs claim?
OpenAI
According to Fox Business, Greg Brockman said: “I think it’s not unreasonable to feel that we are now in the AGI era.” In a TIME interview summarised by The Decoder on 26 August, Sam Altman is reported as saying OpenAI is not there yet but expects an internal system he would call AGI by the end of the year, and that Astra might be the first model to invent new things in a way that matters. The same summary reports Chief Research Officer Mark Chen as estimating the company is “80% of the way” to AGI, and Chief Scientist Jakub Pachocki as saying Astra already meets an internal benchmark of working as an automated research intern. We have read these only through The Decoder, not the original TIME article, and they are OpenAI’s own assessments.
Anthropic
At Davos in January 2026, as reported by Fortune, Dario Amodei predicted that AI would replace the work of all software developers within a year, reach “Nobel-level” research in several fields within two years, and eliminate half of white-collar jobs within five years. These are forecasts, not results. Separately, The Decoder reported on 29 September, citing the Financial Times, that Anthropic’s draft stock-market prospectus warns its technology could pose “existential risks to humanity”. We have not seen the prospectus itself.
Google DeepMind
At the same Davos session, Demis Hassabis said current AI is nowhere near human-level AGI. According to Fortune he put the chance of AGI within the decade at about 50%, estimated five to ten years, and said one or two more breakthroughs are needed, citing gaps such as continual learning, long-term memory, and reasoning and planning.
What do the critics say?
- Yann LeCun, formerly of Meta and now founder of AMI Labs, argued at Davos that large language models will never reach human-like intelligence by training on text alone and that approaches based on “world models” are needed (Fortune).
- Gary Marcus, writing on 3 September, found the ARC-AGI results impressive but not proof of AGI. He called robustness the key open question, said it is disappointing that so little is known about how the system works, and noted that enthusiasts had early access while sceptics did not.
- Toby Walsh (UNSW) told Information Age he would be amazed if Astra had matched all human cognitive abilities and that no single test can capture AGI. Rebecca Johnson (University of Sydney) argued that high benchmark scores cannot by themselves show a system has crossed into a fundamentally different category of intelligence.
- The ARC Prize Foundation said it sees meaningful progress towards generalisation in Astra but is not claiming it is AGI.
Even Altman has called AGI a very poorly defined term, according to the same Information Age article. Walsh also argued that OpenAI needs to reach AGI to keep raising funds. That is his view, and we have not verified any motive.
What does the independent evidence say?
The International AI Safety Report 2026, chaired by Yoshua Bengio and written by more than 100 experts, was published in February, before Astra. As summarised by the law firm Covington, it found that:
| Topic | Finding reported |
|---|---|
| Reliability | Systems still produce false information and are less dependable on multi-step projects; no current method eliminates failures entirely. |
| Agents | They can handle a range of software tasks with limited oversight but lack long-term planning; the report says they complement rather than replace humans. |
| Employment | No significant overall effect found so far; junior workers in exposed roles such as software engineering and customer service may be affected. |
| Loss of control | Expert views on how likely such scenarios are “vary widely”. |
In short, people who study this professionally describe uncertainty, not consensus.
What could this mean for ordinary users?
This section is our analysis, not established fact.
Everyday tools. Assistants that carry out multi-step office tasks are already shipping in limited form. Our guide to which AI assistant suits which task explains how to try them sensibly, and AI in Windows and Office covers what Microsoft offers inside its own software.
Jobs. Forecasts range from dramatic (Amodei’s) to cautious (the Safety Report). A sensible reading is that tasks change faster than whole jobs, and that judgement, domain knowledge and checking work remain valuable.
Security. OpenAI itself rated Astra at its highest cyber threshold and widened access in stages. The practical consequences for you are ordinary: keep systems updated and be wary of convincing fake messages.
Hype. When someone says AGI has arrived or will arrive by a given date, ask who is speaking, how they define it and what they stand to gain.
A simple routine that works either way
- Use AI for first drafts and tedious sorting; keep the final judgement yourself.
- Verify anything with numbers, names or legal consequences.
- Keep your software current and your documents backed up. A supported Windows licence and Office licence keep the everyday basics in order whichever assistant you prefer.
- Revisit the question in six months. For more explainers, see the TheUnitySoft blog.
Frequently asked questions
Has AGI arrived?
Not by any widely accepted definition. OpenAI executives have suggested it is close or arriving; Google DeepMind’s Demis Hassabis and several academics disagree. The answer depends largely on how AGI is defined.
What is the difference between AGI and superintelligence?
AGI is roughly human-level ability across most intellectual work. Superintelligence would go well beyond that. Nobody has demonstrated either.
Should I worry about my job?
The best independent evidence so far, the International AI Safety Report, found no significant overall employment effect, with some risk for junior roles in exposed fields. Forecasts beyond that are opinions. Learning to check and direct AI output is a sensible hedge.
Why do the labs and the critics disagree so much?
Partly because of definitions and partly because of real uncertainty: benchmark results are not the same as reliable real-world performance, and little about how these systems work is public.
Sources
- OpenAI unveils GPT-6 Astra (Fox Business, 3 September 2026). Accessed 21 October 2026.
- OpenAI releasing major upgrade to ChatGPT and Codex with GPT-6 Astra (9to5Mac, 4 September 2026). Accessed 21 October 2026.
- Sam Altman says OpenAI will have AGI by the end of 2026 if you accept his definition (The Decoder, 26 August 2026, summarising TIME). Accessed 21 October 2026.
- OpenAI says the AGI era is here; experts disagree (Information Age, 7 September 2026). Accessed 21 October 2026.
- Hot take on GPT-6 Astra (Gary Marcus, 3 September 2026). Accessed 21 October 2026.
- AI luminaries at Davos clash over how close human-level intelligence really is (Fortune, 23 January 2026). Accessed 21 October 2026.
- International AI Safety Report 2026 examines AI capabilities, risks and safeguards (Inside Privacy (Covington), 12 February 2026). Accessed 21 October 2026.
- Anthropic’s IPO filing shows soaring revenue, mounting costs and existential risks (The Decoder, 29 September 2026). Accessed 21 October 2026.
