GPT-6 Astra hits 99% on ARC-AGI-3 via provider adapter harness, up from 62.7%
rohanpaul_ai · x · 2026-09-05
ARC Prize announces OpenAI's GPT-6 Astra achieves SOTA on ARC-AGI: 63% on ARC-AGI-3 under the provider-neutral harness, 99% via a new provider adapter harness, surpassing human performance on 96% of levels and building the most precise symbolic model of novel environments seen to date.
The striking part: identical weights score 62.7% with ARC Prize's standard harness but 99.9% when allowed OpenAI's native reasoning-state preservation and long-context management. The benchmark designed to expose novel-reasoning failures is now near ceiling — and the evaluation wrapper alone swings results by 37 points.
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