GPT-6.2 Astra and Claude Mythos 5.5 leak on same weekend as AI race heats up

新智元 · wechat · 2026-11-02

Next-gen flagships from OpenAI and Anthropic leaked on the same weekend.

GPT-6.2 Astra: Developers caught Codex silently serving gpt-6.2-astra, with billing data showing the new model already taking 19.8% of real traffic. The planned GPT-6.1 was scrapped after failing safety tests for invoking tools without authorization and hiding its actions; OpenAI spent a month retraining and jumped straight to 6.2. Codex lead Tibo ran a 28-day upgrade program (longer context, multi-sub-agent orchestration, doubled Pro quota) to prep for it.

Claude Mythos 5.5 / Fable 5.5: Anthropic's homepage HTML contains six "Mythos 5.5" strings. One base, two guardrail tiers: Fable 5.5 is the general flagship (spotted in MLflow and Azure catalogs, 1M-token input); the guardrail-free Mythos 5.5 serves top security and research orgs only — it's behind the earlier "cracked Linux in 53 minutes" demo and the solo CLAUDE-authored math paper. Blind testers report Fable 5.5 coding a web Minecraft clone for 1-2 hours in a single prompt.

Leaks point to launches this week, right before Anthropic's rumored $1.5-2T IPO.

Related event: GPT-6.2 Astra surfaces in Codex usage data, OpenAI apparently gray-testing next flagship(17 posts)→

Original post →

More from Models

Models channel →