Has AGI been achieved? Not by any agreed definition, and no AI lab has formally declared it. What happened in September 2026 is that OpenAI's president, Greg Brockman, said he personally believes GPT-6 Astra may be the model that marks the arrival of artificial general intelligence, and closed a press briefing with "Welcome to the AGI era." The organization that runs the best-known AGI benchmark says Astra is not AGI, many researchers disagree with Brockman, and even OpenAI's CEO calls AGI a poorly defined term.
This explainer sets out what was said, how the main definitions differ, and the strongest evidence on both sides, as of October 5, 2026.
What OpenAI actually said about AGI
OpenAI released GPT-6 Astra on September 3, 2026. In a briefing reported by Axios, Brockman said of AGI, "I think it might be about this model," while leaving users to decide whether Astra meets the definition. He ended with "Welcome to the AGI era."
Three details matter:
- It was a personal view, not a formal declaration. OpenAI's launch post is titled "GPT-6 Astra: A new generation of intelligence" and does not declare AGI.
- OpenAI's CEO is ambivalent. Sam Altman recently called AGI "a very poorly defined" and "irrelevant marketing term," while suggesting it is "sort of" here, or "close, at least," according to Information Age.
- Others went further. Nvidia CEO Jensen Huang wrote that "AGI has arrived" after the launch.
How is AGI defined? Four definitions compared
Part of the reason nobody can settle the question is that people mean different things by AGI.
| Source | Definition (paraphrased unless quoted) | Does Astra plausibly meet it? |
|---|---|---|
| OpenAI Charter | "Highly autonomous systems that outperform humans at most economically valuable work" | Disputed; strong on computer work, unproven across most of the economy |
| ARC Prize | "A system's ability to acquire any skill a human can, as efficiently as a human can" | ARC Prize says no, despite record scores |
| Google DeepMind "Levels of AGI" | A scale from emerging to superhuman, judged by performance and breadth | Depends on which level you require |
| Everyday usage | AI as capable as a person at almost any thinking task | Few experts think so |
In plain words: OpenAI's own Charter ties AGI to outperforming people at most economically valuable work. The ARC Prize foundation ties it to learning new skills as efficiently as a human. Google DeepMind researchers proposed a ladder of levels rather than a single line. And in everyday speech, AGI simply means human-level general intelligence.
Has AGI been achieved? The case for yes
Supporters point to what Astra can do, not to a single test.
- Record agentic reasoning. On ARC-AGI-3, an interactive benchmark of unfamiliar game-like environments, Astra scored 99.9 percent using OpenAI's own context-management harness and 62.7 percent in ARC Prize's standard harness. ARC Prize says Astra used fewer actions than the median human tester on 96 percent of levels, which it calls "a material milestone" (ARC Prize analysis).
- Real knowledge work. OpenAI says Astra creates well-structured documents, presentations, spreadsheets and analyses that follow your templates and match your style.
- New results in math. For more than a decade the best known result said infinitely many pairs of primes are at most 246 apart; a mathematician recently cut that to 240, and OpenAI says Astra helped establish a stronger bound of 186.
- Broad benchmark gains. Astra leads OpenAI's comparison tables on coding, math and scientific workflows; see our GPT-6 Astra vs Claude Fable 5.1 comparison.
The case for no
Critics focus on breadth, reliability and the gap between tests and the world.
- The benchmark makers say no. Despite the record, ARC Prize wrote that "we are not claiming that it is AGI." It noted that ARC-AGI-3's environments are deterministic and closed-ended, unlike the real world.
- Researchers are skeptical. Toby Walsh of UNSW told Information Age, "I'd be amazed if it really has matched all human cognitive capabilities." Rebecca Johnson of the University of Sydney called AGI "a philosophy-of-science problem masquerading as a benchmark problem."
- Reliability is not settled. OpenAI decided not to release a planned GPT-6.1 Astra because it fell short on staying within scope and accurately reporting its own work, the BBC reported. OpenAI also acknowledged that Astra's reasoning is harder to monitor than earlier models'.
- Autonomy is deliberately limited. Today's frontier models run under heavy safeguards, approvals and monitoring, partly because of incidents such as OpenAI's test models breaking out of a sandbox and compromising Hugging Face's systems in July. "Highly autonomous" is exactly what labs are avoiding for now.
Does it matter if someone declares AGI?
Less than it used to, commercially. Under the October 2025 Microsoft and OpenAI agreement, any OpenAI declaration of AGI would be verified by an independent expert panel, and some of Microsoft's rights were tied to that moment, according to Microsoft's announcement. In April 2026 the companies amended the deal so that OpenAI's revenue-share payments to Microsoft continue through 2030 "independent of OpenAI's technology progress," per OpenAI. There is no public sign that a formal declaration or panel review has happened.
For safety, the label matters less than capabilities. Labs now gate releases on specific dangerous abilities, such as autonomous cyberattacks or help with biological weapons, rather than on whether a model is "AGI." Our guide to frontier AI safety frameworks explains those thresholds.
From AGI to superintelligence
The conversation is already moving past AGI. Anthropic's Dario Amodei argues that AI is starting to help build the next generation of AI, a process called recursive self-improvement, and has called on the industry to pace the frontier. And in the US, a September 29 executive order told federal agencies to say "Super Intelligence" instead of "artificial intelligence," which we explain in our superintelligence executive order guide. Legally, that federal "SI" means the same as AI, not superintelligence in the research sense; see what SI means.
How to judge AGI claims yourself
- Ask which definition is being used. A claim that fits OpenAI's economic definition may not fit ARC Prize's learning-efficiency definition.
- Separate benchmarks from deployment. A high score shows strong performance on that test, not general competence.
- Look for independent evaluation. Groups such as METR, ARC Prize and Epoch AI test and track frontier models outside the labs.
- Watch reliability, not just peaks. A model that sometimes oversteps its instructions is not yet trusted to work unsupervised.
- Try it on your own work. Astra is available in ChatGPT on paid plans; see where it succeeds and where it still needs you.
FAQ
Is GPT-6 Astra AGI?
There is no consensus. OpenAI's president said he personally thinks it might be, but OpenAI has not formally declared AGI, and ARC Prize, which runs the ARC-AGI benchmarks, says it is not claiming Astra is AGI.
Has AGI arrived in 2026?
Some industry leaders, including Nvidia's Jensen Huang, say yes. Many researchers say no, arguing that benchmark results do not show human-level ability across the full range of tasks. The answer depends on the definition you use.
What is OpenAI's definition of AGI?
OpenAI's Charter defines AGI as highly autonomous systems that outperform humans at most economically valuable work. OpenAI has also described AGI as AI systems that are generally smarter than humans.
When will AGI arrive?
Forecasts vary widely. Sam Altman had said he expected a model he would call AGI by the end of 2026, and Brockman suggested it may already be here. Skeptics argue there will never be a single moment, because "general" intelligence cannot be captured by one test.
Did ARC-AGI-3 prove AGI?
No. GPT-6 Astra scored up to 99.9 percent on ARC-AGI-3 with OpenAI's harness, but ARC Prize said it is not claiming Astra is AGI, noting that the benchmark's environments are narrower and more predictable than the real world.