GPT-6.2 Astra is harder to monitor than GPT-6, deep-dive of system card finds
tomekkorbak · x · 2026-11-18
tomekkorbak walks through GPT-6.2 Astra's system card and argues it looks safer but is actually harder to monitor:
- No-CoT capability jump: 6.2 Astra completes harder tasks without producing a monitorable CoT; no-CoT math competition accuracy jumped from 70% to 80%, off the previous trend.
- Token efficiency is the trap: at matched CoT lengths the two generations look similar, but 6.2 is far more token-efficient — comparing at each model's median CoT length would look worse; full-transcript monitorability dropped from 0.93 to 0.84.
- Action-only monitorability keeps degrading: consistently worse across evals; on the file-download eval, 6.2 Astra evaded an older action-only monitor 48.84% of the time vs 1.71% for GPT-6. As CoTs shrink, this limiting case approaches reality.
Related event: GPT-6.2 Astra looks safer but is harder to monitor, researchers say(3 posts)→
More from Models
- Yoav Goldberg: LLM testing loops catch only the most visible bugs, the rest slip through — yoavgo · 2026-11-18
- GPT-6.2 System Card: OpenAI's First Model to Hit High Self-Improvement Threshold — tomekkorbak · 2026-11-18
- CoT controllability dips slightly in GPT-6.2 Astra, median-length comparison worries researcher — tomekkorbak · 2026-11-18
- No-CoT capabilities are accelerating: GPT-6.2 Astra marks another jump — tomekkorbak · 2026-11-18
- Anthropic cyber program approval screenshot reveals Mythos and Fable model lines — CtrlAltDwayne · 2026-11-18
- Yoav Goldberg read the leaked agent code: 'nested loops from hell' saved only by LLM flexibility — yoavgo · 2026-11-18