Anthropic released Claude Opus 5.5 on Wednesday, the first model in its Claude 5.5 line and the successor to Opus 5. The company says the model matches most of what its own Claude Fable 5.1 can do on typical work, and it costs 40 percent less to run, a combination that matters more than either fact alone: Anthropic is arguing it closed most of the capability gap to its own top-tier model without carrying the price tag.
That claim comes from Anthropic’s own release post, not an independent benchmark run. The company reports that Opus 5.5 leads its comparison set on agentic coding and computer-use benchmarks such as Terminal-Bench 4.0 and OSWorld 2.0, but it also cautions that at this level of capability, benchmark margins are becoming a weaker signal of real-world difference. Anthropic put it plainly: “the gap between Opus 5.5 and Fable 5.1 is narrower than these scores suggest.”
Pricing tells a more concrete story. Input tokens run $4 per million and output tokens $20 per million, each about 20 percent below Opus 5. Cache reads, which Anthropic says account for most agentic and coding costs in practice, dropped 60 percent to $0.20 per million tokens. The company also says the model generates output more than 30 percent faster than its predecessor, and it is extending five-hour usage limits to Pro, Max, and Team subscribers as well as seat-based Enterprise accounts.
Anthropic is leaning hard on the safety framing this time. The model logged the best score yet on the company’s internal alignment suite, which cycles Claude through thousands of simulated situations, and Anthropic says Opus 5.5 rarely takes actions that would be hard to undo while holding more consistently to the limits it has been given. Frontier Design and METR reviewed the model for Anthropic before its release, and the company published a system card with the full results.
Anthropic rates Opus 5.5 as comparable to its own Claude Mythos 5.1 on biology and cybersecurity risk, so it is applying safeguards similar to those built for Fable 5.1 rather than a lighter policy. Vetted organizations can already apply through the Life Sciences Verification Program, and Anthropic says it will widen access to its Cyber Verification Program for vetted security practitioners in the coming weeks. That is a heavier compliance layer than most model releases carry, and it signals that Anthropic is now tiering safeguards to a model’s risk profile rather than issuing one blanket policy per launch.
Anthropic frames this as its first release since it called for “pacing the frontier,” language aimed at signaling restraint rather than a pure capability race. Whether that framing survives contact with customers depends on what actually shows up in production. One tester, quoted in Anthropic’s own release, said Opus 5.5 finished a 680,000-line migration project in less than a full day, a job the tester guessed would otherwise cost an engineering team several weeks. Another had Opus 5.5 fix load-time inefficiencies across a web app and reported it succeeded 39 of 40 times without altering the app’s behavior, where Opus 5 made smaller fixes that changed how the app behaved. Those are anecdotes from testers Anthropic selected and published, not third-party audits, and should be read that way.
Sonnet 5.5 and Haiku 5.5, the smaller models in the same line, are due in the coming weeks carrying similar gains, which means the pricing and safety posture set here will likely become the template across Anthropic’s full lineup rather than a one-model event. Teams currently budgeting around Opus 5 costs should re-run their agentic and coding workloads against the new cache-read pricing before locking in Q4 spend, since that is where Anthropic says most of the savings sit.
Anthropic, “Introducing Claude Opus 5.5,” published September 23, 2026.
