Yoav Goldberg: cheapest run burns ~6-7M tokens per game, mostly reasoning
yoavgo · x · 2026-09-04
AI researcher Yoav Goldberg points out that the cheapest run of a model playing a game consumes roughly 6-7 million tokens per game, with a huge share being reasoning tokens — a substantial cost. He adds an interesting observation: the model was likely trained directly in such an environment, noting that the gap between reasoning and low-reasoning modes on the standard harness is also telling.
More from Models
- Chinese LLM progress can't be chalked up to distillation alone, argue AI commentators — tinyfool · 2026-09-04
- GPT-6 Astra scores 61 vs Fable 5.1's 66 — 'something must be wrong' with the benchmark — Hesamation · 2026-09-04
- Suspected fake 'GPT-6 Astra system card' claims model hides its chain-of-thought — 233C · 2026-09-04
- Rumor: GPT-Astra Access Could Arrive as Early as This Weekend — cyrus_zei · 2026-09-04
- Mystery model "Omen Alpha" speculated to be Xiaomi's MiMo-V3-Flash — teortaxesTex · 2026-09-04
- Claude Fable 5 extended to July 12 — 3 high-leverage ways to use it before access ends — femke_plantinga · 2026-09-04