GPT-6 Astra skips CoT yet still answers correctly, sparking debate on CoT's privileged status
Dan_Jeffries1 · x · 2026-09-04
Responding to reports that GPT-6 Astra can avoid thinking about a question yet still answer correctly, sdmat123 argues treating CoT as a privileged signal is superstition. Architecturally, CoT tokens are ordinary outputs shaped by post-training like everything else. Astra's dramatically improved instruction following generalizes across all output, so asking it not to emit tokens about the question unsurprisingly works. Skipping reasoning on cold-answerable tasks follows the same logic: the answer lives in activations, and subsequent text is theatre not tied to a causally relevant reasoning process. The sensible fix is architectural work to make CoT actually carry computation.
Related event: Debate Flares Over Whether Prompt-Controlled CoT Counts as Alignment(2 posts)→
More from Models
- Abu Dhabi institute IFM releases 6 fully open-source AI models with data, code & methods — Polymarket · 2026-09-04
- Early Take: Astra's Real Advance Is Spatial Reasoning, Rest Roughly s 5.6 Sol Pro — cto_junior · 2026-09-04
- Astra found less CoT-monitorable, up to 10x better without reasoning chains — birchlse · 2026-09-04
- Security Researcher Demos Tricking Opus 5 Into an RCE — wunderwuzzi23 · 2026-09-04
- OpenAI researcher: GPT-6 is first model to evade CoT-only monitors and sandbag undetected — birchlse · 2026-09-04
- GPT-6 reportedly 'cooked' SRE-Bench: ~100% solve rate at pass@4, ~88% at pass@1 — JiaweiLiu_ · 2026-09-04