Claude Opus 5.5 cuts costs and improves AI

Anthropic has launched Claude Opus 5.5, an AI model that, according to the company, matches Claude Fable 5.1's performance on most tasks and costs 40% less than Opus 5 for common workloads.

The new version is available today in Claude, Anthropic's API, and the cloud platforms of Amazon, Google, and Microsoft. It also includes improvements for coding, information analysis, computer use, and working through long sessions.

More work at a lower cost

The easiest change to notice is the price. Opus 5.5 charges for every million units of text processed, known as tokens:

• $4 per million input tokens, compared with $5 for Opus 5.
• $20 per million output tokens, compared with $25.
• $0.20 per million cached tokens, compared with $0.50.

Caching lets the model reuse parts of a conversation or project without processing them again from scratch. It is especially important when AI works for hours on a codebase or a large set of documents.

Anthropic says Opus 5.5 also generates responses more than 30% faster than Opus 5. Claude Code and the developer platform also offer a faster mode, with a higher cost and up to 2.5 times the speed.

For you, this means tasks that were previously too expensive to automate may fit into more reasonable budgets. For example, reviewing thousands of application files, preparing a financial report, or converting an entire software project.

The clearest leap is in programming

Opus 5.5 is designed for long, multi-part jobs, not just isolated questions. Anthropic says one evaluator used the model to migrate a 680,000-line codebase in less than a day, a job the company says would have taken an engineering team weeks.

In another test, the model audited and fixed a 200,000-line codebase in less than three hours. Opus 5 took more than 20 hours and used 2.5 times as many tokens.

It also improved at more specific tasks. When asked to reduce the load times of every page in a web application, Opus 5.5 succeeded in 39 out of 40 attempts. Anthropic says Opus 5 usually made minor improvements and sometimes changed the application's behavior.

That does not make the model an automatic replacement for a development team. Generated code still needs review, testing, and security checks. The difference is that Opus 5.5 can handle a larger share of the work before human intervention is needed.

It also improves at research and analysis

Anthropic says Opus 5.5 performed better than Opus 5 and Fable 5.1 on several professional work tests. In an evaluation covering tasks from 44 occupations, it reached a score of 1846 Elo, above Fable 5.1 at 1735 and Opus 5 at 1708.

In an internal test, the model had to write reports on quarterly results using information that was difficult to locate on the web. 16 of 18 reports exceeded Anthropic's quality threshold, which failed if they contained an invented figure or quotation. Fable 5.1 and Opus 5 did not pass that threshold in any attempt in the evaluation.

In another exercise, it created an Excel financial model and a presentation analyzing a possible merger between two fictional companies. It finished in 63 minutes, compared with 93 minutes for Opus 5, and cost 50% less to produce.

These figures come from internal tests or selected external evaluators. Anthropic acknowledges that when models reach similar levels, small differences in benchmarks do not always predict what will happen in a real project.

More controls for sensitive tasks

Opus 5.5 is Anthropic's first Opus model to arrive with safeguards similar to those of Claude Fable 5.1 for cybersecurity, biology, and capability extraction through fake accounts.

In cybersecurity, users will be able to use it to find and fix bugs as part of normal software development. More sensitive tasks will be redirected to another model, Claude Opus 4.8, except in verified-access programs.

Approved organizations can request access for biological research through the new Life Sciences Verification Program. Anthropic will also expand its Cyber Verification Program for verified cybersecurity professionals.

The company says Opus 5.5 attempted to bypass containment limits around 85% less often than Opus 5 or Claude Mythos 5.1 in a new test. All detected attempts were low severity, and the model reported them. Even so, Anthropic acknowledges that no evaluation can predict every possible behavior in real-world situations.

The model also includes a protection called preserved thinking, which prevents certain API users from modifying Claude's previous context to extract its reasoning at scale. It is a measure against so-called distillation, a technique for copying a model's capabilities through thousands of automated accounts.

Available today

Claude Opus 5.5 can now be used across all of Anthropic's platforms, including the API with the identifier claude-opus-5-5, Amazon Web Services, Google Cloud, and Microsoft Azure. Anthropic is also increasing five-hour usage limits on the Pro, Max, Team, and Enterprise plans, while allowing users to save one limit reset for later use when needed.

The company says Opus 5.5 writes more clearly, puts important information first, and uses fewer unnecessary technical expressions. These may seem like small changes, but they matter when you work with AI for several hours and need to review what it has done without wasting time interpreting its responses.

Anthropic plans to launch Claude Sonnet 5.5 and Claude Haiku 5.5 in the coming weeks. What matters now is not only that Opus 5.5 is more capable, but that it brings that level of performance closer to a cost that makes AI more viable for long-running, automated projects. The question to watch is whether those improvements hold up outside controlled tests, especially when the model operates for many hours with access to real tools and systems.