GPT-6 Astra hits 66% on ARC-AGI-3, near-100% with custom harness at ~$360 per game
AccBalanced · x · 2026-09-05
François Chollet revealed GPT-6 Astra's results on the interactive reasoning benchmark ARC-AGI-3, calling it a step-function change:
- 66% on the standard harness; nearly 100% with a continuous-conversation harness and custom compaction, at roughly $360 per game.
- The continuous-harness version beats the human baseline in action efficiency on almost all levels.
- Inspecting reasoning chains, the team found the model doing efficient on-the-fly symbolic world modeling per game, even inventing its own shorthand DSL—a game-specific algebra.
- ARC's takeaway: capabilities that previously lived in sophisticated harnesses are shifting into the model itself.
A landmark result for interactive reasoning.
More from Models
- Opus 5.5 tops OpenRouter spending in 20 of 29 task categories — mmurph · 2026-10-21
- Meta team builds open-source 30B LLM from scratch in months, hails 'second era' — TimDarcet · 2026-10-21
- DFlash drafter for Muse Glimmer-30B: 5 layers, blockwise 16-token spec decoding — TimDarcet · 2026-10-21
- EvoPress quantization shrinks 30B model from 56GB to 19GB at near-bf16 accuracy — TimDarcet · 2026-10-21
- Radar chart: Muse Glimmer-30B benchmarked against Gemma4-31B and Qwen3.6-27B — TimDarcet · 2026-10-21
- Privacy evals: Muse Glimmer-30B leaks less than Qwen3.6-27B on CIMemories — TimDarcet · 2026-10-21