Anthropic releases Opus 5.5 and cuts pricing by 20%. Your agent calls might secretly get routed to an older model.

Anthropic has shipped Claude Opus 5.5 and made it cheaper to use. The model now costs $4 per million input tokens and $20 per million output tokens, with cache reads at $0.20 and cache writes at $5. Anthropic says the real savings are closer to 40% because the model tends to use fewer tokens and finish output more than 30% faster.

It is also pushing a faster mode for Claude Code and the Claude Platform. That option runs up to 2.5 times faster and is priced at $8 per million input tokens and $40 per million output tokens. On benchmarks, Anthropic says Opus 5.5 lands roughly where Fable 5.1 does on most work, but does better on several agentic coding tests. It scored 66.4% on Terminal-Bench 4.0 versus 55.8% for Fable 5.1, and 54.4% on FrontierCode versus 50.3%.

The bigger story, though, is that the model seems to do the job with less fuss. Box said Opus 5.5 used about a third as many tokens as Opus 5 in its testing while producing answers that were 40% less verbose without losing accuracy. GitHub found it finished more terminal tasks in less than half the steps inside Copilot CLI and VS Code. Deloitte said its lowest-effort setting caught 72% of known bugs in code reviews, compared with 56% for Opus 5 at high effort, and with fewer false alarms.

Anthropic’s own tests point the same way. In one translation task, Opus 5.5 and Fable 5.1 both rewrote HAProxy from C into Rust and passed nearly all regression tests, but Opus 5.5 finished in 9.5 hours instead of 12 and cost 51% less. Another tester audited and fixed a 200,000-line codebase in under three hours, while Opus 5 took over 20 hours and used 2.5 times as many tokens.

But the part builders need to watch is routing. Opus 5.5 ships with the same safety classifiers already used on Fable 5.1, and when one fires Anthropic can transparently send the request to an older model instead. Most flagged cybersecurity requests go to Opus 4.8, while biology and frontier LLM flags go to Opus 5. That means a workflow that looks like it is talking to Opus 5.5 may actually be mixing models mid-stream, which is the sort of thing that makes agent evals look cleaner than production.