OpenAI GPT-6 Astra on ARC-AGI-3

gist

ARC Prize reports GPT-6 Astra as state of the art on ARC-AGI-3 Semi-Private: 62.7% for $26K under the Standard harness and 99.9% for $19K with a Provider Adapter that keeps opaque reasoning state. Astra builds compact algebraic world models, beats the median human action count on 96% of levels, and in PRO-LONG writes game-specific solvers. Saturating ARC-AGI-3 is progress, not proof of AGI.

ideas

• Harness choice moves the scoreboard. Standard notes-only Astra (max) scores 62.7% for about $26K; Provider Adapter with opaque state and compaction reaches 99.9% for about $19K, about 3.66x faster and 49% fewer tokens on shared game-reasoning pairs. ARC Prize will publish both conditions labeled.
• Action efficiency is a new parity line. Against a ~500-person human baseline, Astra (max) used fewer actions than the median completer on 96.0% of levels and 51.7% fewer actions per level on average. Once mechanics click, frontier runs look human-efficient rather than brute-force.
• Symbolic shorthand is the working model. Astra carries strategy notes and invents on-the-fly algebraic DSL for state, plans, controls, and orientation, denser than prior models' notes.
• Tools change the experiment. In PRO-LONG, Astra builds per-game parsers, planners, and libraries such as maze_solver.py; those results are model-plus-tools, not comparable to humans without a code interpreter.
• Saturation is not AGI. ARC-AGI-3 environments are deterministic and closed-ended; ARC Prize treats Astra as a step-function toward generalization while planning harder recursive and open-ended evals.

quotes