Anthropic released Claude Opus 5.5 on September 22 at $4 per million input tokens and $20 per million output tokens; cache reads cost $0.20 per million. The company says a typical workload is about 40% cheaper than Opus 5 and output is more than 30% faster. Fast mode costs twice the token price and runs up to 2.5 times faster. Anthropic’s release and benchmark notes
The useful operating change is cache economics. Long agent jobs repeatedly reread code, instructions and state, so a 60% cut in cache-read price can matter more than the headline token discount. Anthropic also says the model screens each action before execution and is more resistant to prompt injection, but its published benchmark results remain vendor-run or vendor-selected. The deployment question is whether fewer tool calls and shorter outputs survive on an internal workload; Optiver’s early test claims 40–50% lower cost on trading-support tasks, which is promising evidence from a user, not an independent benchmark.