GPT-6 Astra Hits 99.9% on ARC-AGI-3

This is both a model breakthrough and a harness breakthrough: persistent provider-managed reasoning state dramatically changes the result, complicating supposedly apples-to-apples benchmark comparisons.

• –The standard-harness score remains the cleaner cross-provider comparison; the 99.9% result depends on OpenAI-specific context preservation and compaction.
• –Astra reportedly builds compact symbolic world models, tracks mechanics, and plans actions with unusually high information density.
• –Provider Adapter runs were 3.66× faster and used 49% fewer total tokens across shared evaluations.
• –ARC-AGI-3 uses deterministic, closed-ended environments, so benchmark saturation demonstrates strong generalization within scope—not open-ended human-level intelligence.
• –The result makes context management, memory compression, and agent orchestration first-class parts of model evaluation.