Anthropic releases Claude Opus 4.5

Anthropic released Claude Opus 4.5, pitched primarily at coding, agentic and computer-use workloads, at $5 per million input tokens and $25 per million output tokens through the API — a fraction of what Anthropic had charged for its previous flagship, part of a broader fall in frontier-model inference pricing across the industry that year. Anthropic said the model reached state-of-the-art results on SWE-bench Verified and led on seven of eight languages in SWE-bench Multilingual, and reported it outscored Claude Sonnet 4.5 by 29% on its internal Vending-Bench agentic-commerce evaluation and by around 10% on the Aider Polyglot coding benchmark.

A new “effort parameter” let users trade inference cost against performance directly: at medium effort, Anthropic said Opus 4.5 matched Sonnet 4.5’s top SWE-bench score while using 76% fewer output tokens, and cited customer reports of 50–75% reductions in tool-calling and build errors on agentic coding tasks. Anthropic also said the model scored higher than any human candidate on its internal software-engineering qualifying exam, and described it as its most robustly aligned model to date, with improved resistance to prompt-injection attacks relative to prior Claude releases.

The release continued a pattern set by Sonnet 4.5 two months earlier and Opus 4.1 before it: incremental capability gains on coding and agentic benchmarks paired with steep price cuts, as Anthropic and its competitors pushed inference cost down even as raw capability kept climbing, a combination that made frontier-class models cheap enough to run inside high-volume agentic workflows rather than only interactive chat.