When a Lab Declares AGI or an SI Passes a Human-Level Test

Status

Watching As of Nov 8, 2026, no lab has made a formal AGI declaration verified by outside experts, though OpenAI's president said in Sep 2026 the AGI era may have begun.

What AGI means, and who decides

There is no single definition of "artificial general intelligence," or AGI. OpenAI describes it as "highly autonomous systems that outperform humans at most economically valuable work." Researchers quoted by the Australian Computer Society in September 2026 argued that the term lacks a stable definition or measurement, and Sam Altman has called it poorly defined.

For years the most concrete definition sat in a contract. Under the agreement Microsoft and OpenAI announced on October 28, 2025, OpenAI would declare AGI and an independent expert panel would verify it, with some terms ending on verification. On April 27, 2026 the companies changed course: OpenAI's revenue share payments to Microsoft now run through 2030 "independent of OpenAI's technology progress," and Microsoft keeps a non-exclusive license through 2032. Directions on Microsoft reported that this removes the condition tied to AGI. A declaration now carries less contractual weight, though it could still shape policy, investment and public expectations.

The benchmarks behind the claims

Two tests come up most often. Humanity's Last Exam, built by the Center for AI Safety and Scale AI, has 2,500 expert-written questions. Its authors write that high accuracy "would not alone suggest autonomous research capabilities or 'artificial general intelligence.'" ARC-AGI-3, launched in March 2026 by the ARC Prize Foundation, asks systems to learn unfamiliar interactive tasks. At launch every frontier model scored under 1 percent, while human testers solved every environment.

In early September 2026 OpenAI released GPT-6 Astra. OpenAI reported 99.9 percent on ARC-AGI-3 using its own harness. Under ARC Prize's standard harness the model scored 62.7 percent, and ARC Prize said it cleared 96 percent of levels in fewer moves than the median human who solved them. François Chollet said progress was about twice as fast as he expected but did not call it proof of AGI. Brockman said it was "not unreasonable to feel that we are now in the AGI era," while Nvidia's Jensen Huang posted that "AGI has arrived."

Other measures were mixed. ARC Prize called this the first time a model has worked more efficiently than the average human on its tasks, yet the foundation is not claiming Astra is AGI. Epoch AI ranked Astra first on its capability index, while Artificial Analysis placed it level with GPT-5.6 Sol and behind a rival model. Chollet, who had expected ARC-AGI-3 to take about a year to saturate, said his 2030 forecast for AGI had moved sooner.

What changes for people, builders and crypto

For people, a declaration would not change any product overnight. Capabilities arrive through releases, and a label mostly shapes trust, expectations and political debate.

For builders, the Astra results are a lesson in reading benchmark claims. The same model scored 99.9 percent or 62.7 percent depending on the test harness. Compare results run under the same conditions, check whether tools were allowed, and look for independent replication before relying on a headline number.

For crypto, AGI headlines often travel fast through markets and token marketing. A milestone claimed by one lab does not change what a given protocol or agent can do, so on-chain products are best judged by their own audited behavior.

What to watch next

ARC Prize has scheduled ARC-AGI-4 for the first quarter of 2027. Watch whether any lab formally asks an independent expert panel to verify an AGI claim, whether Humanity's Last Exam results approach expert-level scores, and whether lawmakers start citing benchmark results in bills that set capability thresholds. This page will flip to happened when a frontier lab formally declares AGI, or when the creators of a widely used benchmark state that a result meets their bar for AGI.

Signals so far

• Oct 28, 2025: Microsoft and OpenAI agreed that any OpenAI declaration of AGI would be verified by an independent expert panel. OpenAI
• Mar 26, 2026: at the launch of ARC-AGI-3, every frontier model scored under 1 percent, while human testers solved every environment. eWeek
• Apr 27, 2026: Microsoft and OpenAI amended their deal so revenue sharing continues regardless of technology progress, which Directions on Microsoft reported removes the condition tied to AGI. Directions on Microsoft
• Sep 4, 2026: GPT-6 Astra scored 62.7 percent on ARC-AGI-3 under ARC Prize's standard harness, and François Chollet said progress was about twice as fast as he expected. The Decoder
• Sep 7, 2026: OpenAI president Greg Brockman said it was "not unreasonable to feel that we are now in the AGI era," while outside experts disagreed. Information Age (Australian Computer Society)

Questions and answers

Sources

1. The next chapter of the Microsoft-OpenAI partnership OpenAI · Oct 28, 2025
2. The next phase of the Microsoft partnership OpenAI · Apr 27, 2026
3. Microsoft, OpenAI Amend Their Agreement Again Directions on Microsoft · Apr 27, 2026
4. Humanity's Last Exam arXiv (Center for AI Safety and Scale AI) · Jan 24, 2025
5. Every AI Model Fails This New Intelligence Test eWeek · Mar 26, 2026
6. Benchmarks disagree on GPT-6 Astra, but its human-beating efficiency on ARC-AGI-3 pulls Chollet's AGI forecast forward The Decoder · Sep 4, 2026
7. OpenAI says 'the AGI era' is here. Experts disagree Information Age (Australian Computer Society) · Sep 7, 2026
8. Nvidia CEO Jensen Huang declares 'AGI has arrived' after OpenAI unveils GPT-6 Astra Fox Business · Sep 8, 2026