The creator of the most prominent AGI benchmark believes that AGI is still a few years away.
François Chollet, co-founder of the ARC Prize and the researcher behind the ARC-AGI benchmark — widely regarded as the most rigorous test of progress toward artificial general intelligence — has put a rough date on when he expects AGI to arrive: the early 2030s.
Speaking recently, Chollet said: “The timeline to AGI — if you just try to extrapolate from the current rate of progress and the amount of investment that’s going into not just the LLM stack, but also side ideas and side bets that might work out — I think we’re probably looking at AGI by 2030, early 2030s, most likely.”
He then offered a telling frame of reference: “Around that time, we’re going to be releasing maybe ARC-AGI 6 or ARC-AGI 7. That’s probably going to be AGI.”
The ARC Prize has been releasing newer and newer versions of its benchmark as older ones get saturated by AI models. ARC-AGI-3, released on March 24, 2026, shifts from static visual puzzles to interactive turn-based environments where an agent must figure out goals, rules, and strategy with no instructions. Humans solve 100% of the environments; the best AI model scores 0.37%. ARC-AGI-1 and ARC-AGI-2, by contrast, are effectively saturated — Gemini 3 Deep Think scored 84.6% on ARC-AGI-2 and 96% on ARC-AGI-1. Each new version of the benchmark raises the ceiling precisely to prevent AI systems from appearing more capable than they are.
Chollet’s early 2030s timeline puts him closer to the cautious end of the spectrum. Several AI lab CEOs have predicted AGI within the next year or two. But Chollet has been a consistent sceptic of the LLM path to AGI — he has argued that LLMs are fundamentally bottlenecked by human-generated data, not compute, and that true AGI would scale with compute alone, unconstrained by the availability of human output. His timeline, notably, accounts for “side bets” beyond the LLM stack — alternative approaches that could compress or extend the path depending on what breaks through.
At the current pace, with a new, harder ARC-AGI benchmark dropping roughly every year, Chollet’s prediction implies six or seven more generations of increasingly difficult tests before the bar is finally cleared. And whether the field gets there by scaling existing architectures or through something fundamentally different remains the central open question in AI research.