Epoch AI: GPT-6.2 Astra performs 12.3 serial reasoning steps in a single token, no CoT needed
morqon · x · 2026-11-18
Epoch AI measured how many serial reasoning steps models can perform in a single output token without chain of thought: GPT-6.2 Astra manages 12.3 steps, beating GPT-6 Astra (10.5) and GPT-6.1 Sol (8.6). Methodology: steps counted at 25%-above-chance accuracy, complexity-normalized.
Historical context: before GPT-6 Astra, the best was Claude Fable 5.1 at 4.2 steps; the oldest model assessed, Claude Opus 3, managed only 2.3 — a huge leap across a few generations. Reposter AlexBarry4 adds:
- No-CoT measurements reveal the true scale of Astra's leap; other approaches may understate it.
- Fable 5.2's knowledge-benchmark performance suggests it is either very large or uses Astra-like architectural changes.
Why it matters: models capable of complex reasoning without CoT are harder to monitor — directly relevant to safety evaluation. Full paper coming soon.
Related event: Epoch AI measures GPT-6.2 Astra at 12.3 serial reasoning steps per token(2 posts)→
More from Models
- Astra 6.2 plays detective: finds URL on a projector screen in hours of B-roll — gabrielchua · 2026-11-18
- Qwen 3.8 27B vs Flash-Next vs DeepSeek V4.1 Flash: one-shot landing page build-off — jesdga95 · 2026-11-18
- GPT-6.2 Astra plays RimWorld for 250 in-game days and builds a decent colony — jxnlco · 2026-11-18
- Polymarket claims GPT-6 Cyber uncovered 35 security flaws across Apple iPhones and Macs — Polymarket · 2026-11-18
- OpenAI ships GPT-6.2 Astra: 22-point ExploitBench jump, tighter cyber safeguards — charliermarsh · 2026-11-18
- Halocline inference engine hits 148-182 tok/s on Qwen-Flash-Next with KLD 0.082 — Distinct-Pie2389 · 2026-11-18