Running a model at 1.5 min per move: Mike Frank tunes reasoning-time settings over OpenRouter
MikePFrank · x · 2026-09-15
Mike Frank shared details of tuning a model's reasoning for a game: currently set to Max reasoning with a 1.5-minute target per move against a 1.5-hour game clock (first 40 moves), which would use about an hour of thinking. He's considering dialing back to High or 1 minute per move—OpenRouter providers run this model at roughly 100 toks/sec, so a minute is 6,000 tokens, which he suspects the model could easily use. The model's inherent verbosity also contributes to slowness.
More from Models
- Twin prime bound pushed to 186 as GPT-6 Astra launch fuels lab math race — RexDouglass · 2026-09-15
- OpenAI cuts desktop voice pricing ~60%, 2.4x more ChatGPT Voice in Codex — athyuttamre · 2026-09-15
- Lobehub bench shows high reasoning beats max, blogger slams vendor marketing spin — karminski3 · 2026-09-15
- Qwen 27B Passes the Pelican Test for $0.21 vs $3 on Paid Models — DivideHorror3217 · 2026-09-15
- LlamaIndex hits back: Cohere Parse scores only ~50% on ParseBench despite $1.50/1k pages pricing — ravithejads · 2026-09-15
- DeepSeek engineer who wrote v4.1's core Attention operator reflects on building his own replacement — WebAssemblyMan · 2026-09-15