Counterintuitive: Opus 5.5 fails 2x more turns on max than on high, in every time bucket
tristanbob · x · 2026-11-08
A test by AdamHoltererer found Opus 5.5 fails roughly twice as many T3 turns on max reasoning effort as on high. It's not a timeout artifact: max fails more in every turn-length bucket, and most extra failures occur within the first 30 seconds. Sol shows the same pattern at smaller scale (1.4x). Cause unknown; tristanbob's advice: just use high.
More from Models
- September's model release blitz makes a packed October feel dead — haider1 · 2026-11-08
- Token Limit Resets Spark Debate: Should Employers Cover Employees' AI Costs? — dan_s_becker · 2026-11-08
- Anthropic analyzed 207K research sessions: scientists lean on Claude for data analysis but barely use it for ideation — dotey · 2026-11-08
- Qwen-4-Flash-Early tops GameFeel Bench at 49.9, ~7.6x cheaper than rivals — pbaylies · 2026-11-08
- Benchmarks get gamed: vendors sell solutions to labs, so 'emergence' is mostly training — sebkrier · 2026-11-08
- llama.cpp adds support for Mage-VL-Cowork vision model via new mtmd PR — pmttyji · 2026-11-08