GPT-6 Astra uses reasoning tokens in only 27% of ARC-AGI-3 actions, hinting at a leap in opaque reasoning
mhmazur · x · 2026-09-05
New data point from ARC-AGI-3 on GPT-6 Astra
Under low reasoning effort in a standard harness:
- GPT-5.6 Sol used reasoning tokens in 98% of its actions while playing ARC-AGI-3 public games
- GPT-6 Astra used them in only 27% of actions (!)
Semi-private games showed the same pattern, suggesting the difference isn't driven by public benchmark contamination and may instead reflect architectural changes in how the GPT models work.
In the surrounding discussion, the shift is framed as a potentially massive jump in opaque reasoning: GPT-6 Astra appears able to solve hard competition math problems entirely "in its head" without verbalized reasoning, where prior models could only handle basic word problems. The author flags this as concerning if the benchmark results are representative, while noting the highlighted caveats. Still third-party analysis pending further validation.
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