1,819 rollouts show GPT-5.6's thinking budget quietly dropped 24% since Sept 4
tensorqt · x · 2026-11-21
To test whether models secretly get nerfed, the authors ran the same RL environment 1,819 times from August to November — 39,926 turns and 1.5B tokens — tracking thinking, turns, and speed.
Findings: OpenAI appears to dynamically adjust the thinking budget ("juice") at fixed effort settings — GPT-5.6 Luna's thinking dropped 24% from Sept 4 with identical scores and speed; GPT-6 Astra dropped 13% in November. Users get less reasoning at the same tier while headline metrics stay flat.
More from Models
- Mathematician and Claude claim a proof of the Riemann Hypothesis — Kyrannio · 2026-11-21
- GPT 6.2 Astra scores 41.2% on robotics benchmark Ember-0, more than doubling since September — k7agar · 2026-11-21
- GPT-6.2 Astra review: stronger at math and economic modeling, now with Ultrafast — soumitrashukla9 · 2026-11-21
- Google launches first interactive Gemma 4 explainer teaching how LLMs work in the browser — DynamicWebPaige · 2026-11-21
- Persimmon built a safety framework from scratch for human-like user models — gharik · 2026-11-21
- InferenceBench recap: Anthropic led for half a year before Google's dramatic comeback — maksym_andr · 2026-11-21