Benchmarks usually fade slowly. A new test appears, scores creep up over years, and eventually it stops telling anyone anything. ARC-AGI-3 did not get years.
- 0.51%frontier AI at launch, 25 March 2026
- 62.7%GPT-6 Astra with ARC Prize's standard harness, 3 September 2026
- 99.9%the same model with a harness that keeps its reasoning state
March
ARC Prize launched ARC-AGI-3 on 25 March 2026: a set of interactive game environments rather than static puzzles. The launch result was stark. In ARC Prize's own words: “Humans score 100%. Frontier AI scores 0.51%.”
September
On 3 September, ARC Prize published its results for OpenAI's GPT-6 Astra on the semi-private set. With the standard harness, in which the model carries its own notes from step to step, Astra scored 62.7%. With a harness that preserves the model's reasoning state between steps, it scored 99.9%.
ARC Prize also measured efficiency, not only success. Its human baseline is the median number of actions that about 500 members of the public needed to complete each level. On 96.0% of levels, Astra at its highest setting needed fewer.
“A 100% score means AI agents can beat every game as efficiently as humans.”
From half a per cent to well over half, in 162 days.
The harness matters
The gap between 62.7% and 99.9% is the same model with different scaffolding: the software that decides what the model remembers from one step to the next. ARC Prize will report both results from now on, each labelled.
We give both here too. Even the standard figure alone is a move from under one per cent to well over half in six months.
What to watch
It is the pattern we described in The length of a task: curves that look gradual from a distance are steep up close. A benchmark is useful only while it can still be failed. The question now is what comes after ARC-AGI-3, and how long the next one lasts.