GoKawiilOpenAI's GPT-6 Astra hits 99.9% on ARC-AGI-3 with provider adapter harness
OpenAI's unreleased GPT-6 Astra model scored 62.7% on the ARC-AGI-3 Semi-Private benchmark using a Standard harness for $26,000, and 99.9% using a Provider Adapter harness that preserves reasoning state between requests for $19,000. The model also beat the median human tester's action efficiency on 96% of the benchmark's levels, building compact symbolic world models and its own shorthand notation to track game states.
GoKawiil's interpretation of the reporting above, not reported fact.
ARC-AGI-3 is designed to measure how close AI systems are to human-level general intelligence by testing exploration and planning in novel environments without instructions. Near-perfect performance under a harness that lets the model retain internal reasoning state suggests significant progress in agentic reasoning, though the huge cost gap between harness types highlights how much technique and infrastructure choices still shape reported benchmark results.
- GPT-6 Astra reached 99.9% on ARC-AGI-3 with a Provider Adapter harness versus 62.7% with a Standard harness.
- The model outperformed median human testers in action efficiency on 96% of benchmark levels.
- It built symbolic world models and custom shorthand to represent game rules and plan actions.
Source: arcprize.org, 2026-09-03
Published there as: “OpenAI's GPT-6 Astra on ARC-AGI-3”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.