Chollet Moves Up His AGI Timeline Again After GPT-6 Astra's ARC-AGI-3 Score

François Chollet, creator of the ARC-AGI benchmark and a longtime skeptic of near-term AGI and LLM scaling, moved his AGI timeline up again to before 2030 after seeing GPT-6 Astra's ARC-AGI-3 results. The statement follows the fastest-ever saturation of the benchmark, and is notable precisely because it comes from one of the most prominent AGI skeptics.

Confirmed

• Chollet's timeline keeps shrinking: in 2024 he estimated AGI was about 10 years away; by February 2026 he had shortened it to around 2030; after GPT-6 Astra's ARC-AGI-3 score he now says "earlier" (traced by @haider1 and @relegi)
• He said that when ARC-3 launched he expected frontier models to take about a year to saturate it; GPT-6 Astra's progress reached saturation in only 6 months, roughly twice as fast as expected (@fchollet)
• Chollet clarified that even saturating ARC-3 does not constitute proof of AGI, that the ARC team makes no such claim, and that their knowledge of the system is limited to benchmark scores (@fchollet)
• He views benchmarks and models as co-evolving: new benchmarks challenge models with novel problems, shape research directions, and are adjusted as models improve, targeting the residual gap between AI and human intelligence (@fchollet)
• ARC roadmap: ARC-4 is in development, planned for Q1 2027; ARC-5 is planned; the final version will likely be ARC-6 or 7 (@fchollet, relayed by @MartinSignoux)
• In replies, Chollet predicted AI performing typical human computer tasks will arrive sooner than expected; @Afinetheorem reads this as him joining the "by 2029 at the latest, AI can do anything an average person does on a computer" camp
• Reviewer @scaling01 also voiced unease: he expected the new model to score 70%+ on ARC-AGI-3, yet overall capability came in about 2.5 ECI points above his expectations despite his already-bullish stance; this kind of overshoot happened once last year too, and he finds the current race-like release sprint unsettling

Why it matters

• As the ARC Prize author and one of the most influential skeptical voices on AGI timelines, Chollet's shifting stance is often read as a signal of frontier capability progress
• ARC-3 saturated at roughly twice the expected pace, suggesting targeted optimization and model progress jointly accelerated the benchmark's defeat—and explaining his motivation to fast-track ARC-4
• His repeated insistence that benchmark scores do not equal AGI offers a calibration frame from the benchmark's own author for interpreting the new wave of high-score hype, while outside reviewers' concerns confirm that this capability jump exceeded expectations across the board.

2026-09-04 ~ 2026-09-04 · 8 related posts

• [source] Chollet: Saturating ARC 3 would not prove AGI — it was never meant to be a finish line — fchollet · 2026-09-04
• [source] Chollet: ARC-AGI-4 lands Q1 2027, and solving ARC-3 is not AGI — fchollet · 2026-09-04
• [source] Chollet: Astra Saturated ARC-AGI-3 Twice as Fast as I Predicted — fchollet · 2026-09-04
• Fchollet: ARC-4 due early 2027, final ARC likely 6-7, AGI around 2030 — MartinSignoux · 2026-09-04
• François Chollet pulls AGI timeline to sooner than 2030 after Astra and ARC-AGI-3 — relegi · 2026-09-04
• Chollet says computer-use AI will arrive sooner; insiders whisper 2027 — Afinetheorem · 2026-09-04
• After GPT-6 Astra's ARC-AGI-3 score, Chollet says AGI is coming 'sooner' — haider1 · 2026-09-04
• Model capabilities keep beating bullish forecasts, sparking 'reckless race' fears — scaling01 · 2026-09-04