Paper: Max reasoning effort costs up to 42x more and often fails tasks lower effort solves
zainhas · x · 2026-10-29
An arXiv paper measures the counterfactual effect of reasoning-effort levels on proprietary and open-weight models across four benchmarks, spanning single-call reasoning and long-horizon agentic tasks.
- Higher effort isn't monotonic: it can raise aggregate accuracy but frequently fails on tasks that lower effort solved.
- Wasteful on easy tasks: on tasks every effort level solves, the highest effort costs up to 42x the lowest.
- Effort changes behavior: at higher effort, agents verify their own work more and search more instead of working through problems directly.
The authors frame reasoning effort as a credence good: buyers see the outcome but can't assess whether the compute they paid for was necessary.
Related event: Max reasoning effort costs 42x more, often unnecessary(4 posts)→
More from Models
- Early Take on Gemini 4 Carbon: Lots of Thinking, Good Model Overall — SamuelAlbanie · 2026-10-29
- Gemini 4 Carbon tops AA Index v5, beats Opus 5.5 and GPT-6 Astra on most benchmarks — Yuchenj_UW · 2026-10-29
- Gemini 4 Carbon Keeps $2/$10 Pricing, Scores 59 on AA v5.0 — _philschmid · 2026-10-29
- Reddit thread: picking third-party LLM inference providers, GLM Flash cited as best value — inwardPersecution · 2026-10-29
- Reddit reposts Gemini 4 Carbon announcement with benchmark table — Marimo188 · 2026-10-29
- Google Announces Gemini 4 Carbon, Its Next Frontier Model — patloeber · 2026-10-29