Anthropic launches Opus 5.5 with cost cuts and benchmark gains

Anthropic unveiled Opus 5.5 on September 23 2026. The model targets Claude Fable 5.1 in performance while cutting usage costs nearly in half. It is priced at four dollars per input token and twenty dollars per output token per million tokens. Cache read costs drop to zero point two dollars and cache write costs fall to five dollars. Opus 5.5 uses fewer tokens for long tasks and runs up to two point five times faster in fast mode priced at eight dollars per input token and forty dollars per output token per million tokens.

Subscribers gain extended rate limit resets across Pro Max Team and Enterprise plans. The model excels at large code migrations such as moving a two hundred thousand line code base from C to Rust in nine point five hours compared to twelve hours for Fable 5.1. It also completes thirty nine of forty web page load speed tests where Fable 5.1 shows limited improvement. Opus 5.5 matches GPT 6 Astra on Terminal Bench 4.0 with lower cost and outperforms GPT 5.6 Sol on Cursor Bench 11 points with one third the expense.

It scores higher than competitors on Knowledge Work benchmarks including GDPval AA v2.1 OSWorld 2.0 and Humanity’s Last Exam. The model passes sixteen of eighteen financial report checks where earlier models fail entirely. Gray Swan testing shows Opus 5.5 resists prompt injection as well as Fable 5.1. The company adds a classifier sandbox and open source safety checks to support long autonomous agent runs.

Reported by Techsauce.