AGI
Definition of done
This arc is done when an AI system that the public can use, or that outside testers have had full access to, is shown by independent evaluation to do most remote knowledge work at the level of a skilled professional across many fields. Concretely: at least one published evaluation run by someone other than the developer, covering a broad range of occupations, finds the system matches or beats expert professionals on a majority of tasks, and an independent autonomy test shows it can carry long, multi-day projects to completion with little human help, at a cost that makes it a real substitute. It must also work the way an employee does: given a goal at the level a manager would assign, it plans, starts and drives the work to completion on its own, asking only when it is actually blocked, and it keeps what it learns on the job, from feedback and its own mistakes, so it improves over weeks without retraining. It does not count if a company declares it has reached AGI or superintelligence, if a government or company renames AI, if one benchmark (such as ARC-AGI) is saturated, if a single striking result (such as a math proof) lands, or if the only evidence is vendor-run evals and demos. Physical and robotic work is not required.
Where it stands
No system meets this bar as of late October 2026, and no independent evaluation has claimed one does. The frontier is split three ways: OpenAI's GPT-6 Astra (September 3), Anthropic's Claude Opus 5.5 (September 22) and Google's Gemini 4 Argon (September 30, widely available from October 13) each lead different indexes, and METR's pre-release testing of Opus 5.5 judged it unlikely to fully automate AI research. Argon's verified 97.5% on ARC-AGI-3 drops to 49.8% under the standard harness, and OpenAI's large batch of claimed math results still has no independent expert reading. The US government's September 29 order to call AI "super intelligence" is a renaming and does not move this arc.
Milestones
2026-10-15: ARC Prize verifies Gemini 4 Argon at 97.5% on ARC-AGI-3 with Google's own adapter harness but 49.8% with the standard one, making Google DeepMind the second lab above 95% — ai — Gemini 4: the Kavukcuoglu ship window
2026-10-06: OpenAI publishes results on open math problems from an unreleased internal model, a batch The Verge counted at 722 manuscripts, released alongside Lean proof formalizations, after its September 8 Navier–Stokes claim, though no independent expert has yet read the batch — ai — OpenAI Navier–Stokes claim + same-day attribution fight
2026-09-30: Google ships Gemini 4 Argon, first to trusted cyber defenders, claiming first place on the Vals Index and Text Arena, with broad developer and subscriber access following on October 13 — ai — Gemini 4: the Kavukcuoglu ship window
2026-09-22: Anthropic releases Claude Opus 5.5, which tops the independent Artificial Analysis Intelligence Index, while METR's pre-release testing finds only a modest gain on AI research work and judges full automation of it unlikely — ai — Anthropic post-pacing model line: Opus 5.5
2026-09-03: OpenAI launches GPT-6 Astra two days after Anthropic's Claude Fable 5.1, and independent indexes split the pair, with Artificial Analysis putting Fable ahead on intelligence and the two tied on coding — ai — GPT-6 Astra rollout
Related claims
- OpenAI — capabilities sufficient to fully automate AI research within 18 months, as relayed by board member Paul Christiano (said 2026-09-09; due 2028-03-09, review by 2027-09-09): pending.
- Anthropic — an embedded outside evaluation team with employee-like access and freedom to publish, in the near future (said 2026-09-12; review by 2026-12-15): pending.
- White House and frontier labs — Accord on Superintelligence, a voluntary joint commitment on frontier responsibilities with internal and external review (said 2026-09-29; review by 2026-10-29): pending.
- Sanders and Casar — a forthcoming Ban Artificial Superintelligence Act to permanently ban superintelligent AI and pause advanced AI development until a new regulator sets rules (said 2026-09-04; review by 2026-10-15): pending.