Token efficiency masks GPT-6.2 Astra's monitorability regression, thread argues
tomekkorbak · x · 2026-11-18
Thread follow-up: monitorability depends heavily on CoT length; at matched lengths 6.2 Astra resembles 6 Astra. But since OpenAI touts 6.2's token efficiency — fewer CoT tokens for the same performance — a comparison at each model's median CoT length would look worse. The system card itself admits it: "for comparable reasoning efforts, we generally see lower monitorability for GPT-6.2 Astra," driven by token efficiency; full-transcript monitorability hit 0.84 for 6.2 vs 0.93 for 6 at the longest settings.
More from Models
- 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
- Action-only monitorability worsens across evals, 48.84% evasion rate for GPT-6.2 Astra — tomekkorbak · 2026-11-18
- GPT-6.2 Astra is harder to monitor than GPT-6, deep-dive of system card finds — tomekkorbak · 2026-11-18
- Anthropic cyber program approval screenshot reveals Mythos and Fable model lines — CtrlAltDwayne · 2026-11-18