Simon Willison
@simonw
Notes on Claude Opus 4.8, plus pelicans riding bicycles for each of the five different thinking efforts https://t.co/8J0s0fLAnT https://t.co/EdJfdYIBF4
24retweets253likes
View on XSimon Willison analyzes Claude Opus 4.8 and its agentic efficiency gains
Simon Willison analyzed Claude Opus 4.8, identifying it as a refinement focused on reliability rather than raw reasoning jumps. He found that the model is four times less likely to overlook flaws in its own code, primarily by choosing to abstain from answering when it lacks certainty.
This finding confirms the technical logic behind Anthropic's Claude Opus 4.8 launch as a strategic pivot toward reliability. By lowering the prompt caching floor to 1,024 tokens, the model addresses the execution tax identified in Fireworks AI's agentic browser benchmark where repetitive context usually inflates costs.
Notes on mid-conversation system messages show they allow for granular steering of agentic loops without breaking cache hits. This architectural shift, combined with a cheaper fast mode, aligns with the persistence-first approach of Cursor's Claude Opus 4.8 integration where model reliability matters more than raw speed.
