Opus 5.5 botches a simple experiment: picks a constant-token task, then blames the API
adam_karvonen · x · 2026-11-19
Researcher Adam Karvonen shares a funny Opus 5.5 story: asked to check token usage vs. reasoning effort, the model picked a counting task (which obviously needs a constant number of tokens), cycled through a pile of API parameters, and concluded that effort wasn't supported.
More from Fun
- OpenAdvisor launches as 'Yelp for PhD advisors' with anonymous reviews and rebuttals — Pseudomanifold · 2026-11-19
- Pedro Domingos: Anthropic suffers from FOOM FOMO — pmddomingos · 2026-11-19
- 'Taste is the only moat' — VCs right before funding gambling apps and 12,000-fake-star startups — vasuman · 2026-11-19
- Founder: I never understood EAs until I had to coordinate 8 meetings at once — gabriel1 · 2026-11-19
- Kevin Roose reveals Anthropic's near-names: Squirrel, Unfoom and Endosymbiosis almost won — kevinroose · 2026-11-19
- Fake prompt with empty <think> block shows models saying Yes without reasoning — Sauers_ · 2026-11-19