OpenAI's GPT-6 Astra on ARC-AGI-3 | ARC Prize
Summary
OpenAI's GPT-6 Astra achieves state-of-the-art results on ARC-AGI-3, scoring 99.9% using a Provider Adapter harness and surpassing human action efficiency on 96% of levels.
Astra surpasses human action efficiency.
In the Provider Adapter harness, Astra used 51.7% fewer actions per level on average than the median human participant.
Astra (max) used fewer actions than the human baseline on 96.0% of levels
Astra builds compact symbolic models.
The model generates an on-the-fly algebraic shorthand to track game state, coordinates, and multi-step plans in a dense, code-like notation.
It distilled the scene into a compact code-like symbolic model
Provider Adapters boost performance.
Using provider-designed context management increased Astra's Semi-Private score from 62.7% to 99.9% and reduced total token usage by 49%.
Provider Adapter runs were approximately 3.66x faster
Astra develops custom software tools.
In the PRO-LONG harness, the model created game-specific libraries including board parsers, search algorithms, and Python files like maze_solver.py.
Astra even produced small, game-specific software libraries
Saturating ARC-AGI-3 is not AGI.
The benchmark has a tightly bounded scope with deterministic mechanics and does not represent the open-ended complexity of the real world.
It does not represent the complexity and open-endedness of the real world
Read the original ↗Save to your Sponge
Summarized with Sponge. Quotes are the author’s own words.