Epoch: GPT-6.2 Astra performs 12.3 serial reasoning steps in a single output token
scaling01 · x · 2026-11-18
- Epoch AI measured how many serial reasoning steps models can perform within a single output token, without chain of thought (steps normalized for complexity, measured at 25% above-chance accuracy).
- GPT-6.2 Astra hits 12.3 steps, versus 10.5 for GPT-6 Astra and 8.6 for GPT-6.1 Sol; the previous best was Claude Fable 5.1 at 4.2, and Claude Opus 3 managed only 2.3.
- Epoch notes models capable of complex reasoning without CoT may be harder to monitor; scaling01 jokes that "loopmaxxers are the next paradigm."
More from Models
- Early GPT-6.2 reviews split: Matt Shumer calls Astra/Sol essentially unusable — Independent-Wind4462 · 2026-11-18
- Epoch AI: GPT-6.2 Astra scores 174 on ECI, matching Fable 5.2 — scaling01 · 2026-11-18
- Gemini 4 Carbon vs GPT-6.2 Astra head-to-head on an SVG demo — Angaisb_ · 2026-11-18
- Microsoft's MAI Models Go Live on Copilot, OpenRouter and Vercel — marlene_zw · 2026-11-18
- Theo's Long Critique: GPT-6.2 Astra Is Painful to Use, Opus 5.5 Still Better Overall — Angaisb_ · 2026-11-18
- Microsoft announces MAI Decision 1, Code 2, Realtime 1 and more at Ignite — testingcatalog · 2026-11-18