Anthropic launched Claude Opus 5.5 on September 22, 2026, describing it as the first model in its Claude 5.5 family. The release targets demanding coding, computer-use, and knowledge-work tasks while reducing the token rates and typical workload costs compared with Opus 5.
Anthropic lists API pricing of $4 per million input tokens and $20 per million output tokens. It also reports stronger coding and agent performance, faster output, and improved results on selected safety evaluations. These are company-reported results; teams should test Opus 5.5 on their own work before switching production systems.
Last verified: September 22, 2026, using Anthropic’s announcement and system card. Platform access and model safeguards can vary by provider and account.
Claude Opus 5.5 at a glance
- Model ID:
claude-opus-5-5 - Input price: $4 per million tokens
- Output price: $20 per million tokens
- Cache reads: $0.20 per million tokens
- Cache writes: $5 per million tokens
- Availability: Anthropic says it is available across Claude platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure.
- Positioning: coding agents, computer use, long-running professional tasks, and knowledge work.
Anthropic says Opus 5.5 costs 40% less to run than Opus 5 on typical workloads at default settings and produces output more than 30% faster. That workload estimate reflects Anthropic’s tests and usage patterns; actual results depend on prompt size, reasoning settings, caching, tools, and how much correction a task needs.
What changed in Opus 5.5
More capable agentic coding
Anthropic highlights codebase-wide migrations, software audits, debugging, and multi-step terminal work. Its announcement includes results from Terminal-Bench, FrontierCode, and CursorBench, alongside examples from early testers. The company also notes that benchmark gaps are becoming less predictive of real-world differences and that its comparison conditions vary by test.
For a development team, the useful question is whether the model can finish a representative task with fewer steps, fewer mistakes, and less review time. Test on a safe branch, set a clear acceptance checklist, and track whether the model’s changes pass the project’s normal tests and review process.
Lower token prices and better cache economics
| Token type | Opus 5.5 price per 1M tokens | Opus 5 price per 1M tokens |
|---|---|---|
| Input | $4.00 | $5.00 |
| Cache writes | $5.00 | $6.25 |
| Cache reads | $0.20 | $0.50 |
| Output | $20.00 | $25.00 |
Anthropic says the model’s lower serving cost, lower token prices, and reduced tokens per task combine to cut typical workload costs by 40% compared with Opus 5. The announcement also says subscription limits on several plans are increasing. This does not guarantee a 40% saving for every workflow: measure your own cost per successful outcome, including tool calls and retries.
Improved computer use and professional work
Anthropic reports gains on computer-use and professional-work evaluations, including an internal financial-analysis exercise and external evaluation results. Such tasks can involve research, document analysis, software changes, and multi-step tool use. They are also sensitive to data quality and context, so verify cited facts, calculations, and actions before relying on results.
Clearer communication during longer tasks
Anthropic says Opus 5.5 brings key information forward, follows writing instructions more consistently, and uses less jargon than its predecessor. This can make long-running agent work easier for a person to follow and review. Communication improvements do not guarantee that an explanation is complete or that the work itself is correct.
Safety and deployment controls
Anthropic reports that Opus 5.5 performed strongly on its automated behavioral audit and improved on several measured behaviors related to misleading claims, sandbox boundaries, and harmful actions. It also says the model attempted to circumvent containment boundaries less often than Opus 5 in a specific evaluation. These are controlled tests, not a promise that the model will never make an unsafe choice in a new environment.
The release includes safeguards for cybersecurity, biology, and model distillation. Anthropic says routine software-development bug finding and fixing remain available, while most cybersecurity tasks are routed to another model; verified researchers can apply for more permissive access programs. The announcement also notes that some safety evaluations remain difficult to generalize to real-world use.
For production systems, keep tool access narrow, isolate code execution, log model actions, and require human approval before external messages, production changes, sensitive-data operations, or other consequential actions. Treat model-generated code as a proposed change that still needs ordinary review and tests.
Availability and API integration
Anthropic says Opus 5.5 is available across its platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure. The Claude Platform model ID is claude-opus-5-5. Exact rollout, region, account, and deployment options can differ by provider.
The API launch also includes preserved thinking, Anthropic’s safeguard intended to prevent API users from editing prior context to extract model reasoning. Anthropic says it applies to Opus 5.5 and Fable 5.1 for API accounts created on or after August 31, 2026. Existing integrations that depend on editing prior thinking content should review the official preserved-thinking documentation and update their handling accordingly.
How teams should evaluate the release
Start with a small set of real tasks and compare Opus 5.5 with the model you use today. Record completion quality, factual errors, latency, total token and tool cost, retries, and how much human correction was needed. Include difficult examples and failures, not just tasks that make the new model look good.
For tool-using agents, separately test permission boundaries, prompt-injection resistance, rollback behavior, and what happens when a tool returns incomplete or misleading information. Anthropic’s own evaluations are useful starting points, but your data, instructions, tools, and risk profile determine whether the model fits your application.
How I can help you apply Claude Opus 5.5
I provide AI consulting and custom AI development for teams choosing models, integrating APIs, building evaluations, and setting up safe agent workflows. I can also help with SaaS product engineering, workflow automation, and AI features for websites and mobile apps.
The best first step is usually a controlled evaluation of one valuable workflow. That gives you a real comparison of quality, speed, cost, and review effort before you change a larger system.
Official sources
Frequently asked questions
What is Claude Opus 5.5?
Claude Opus 5.5 is Anthropic's new Opus model and the first release in its Claude 5.5 family. Anthropic positions it for demanding coding, computer-use, and knowledge-work tasks, with lower API token prices than Opus 5.
What is the Claude Opus 5.5 API model ID?
Anthropic lists the Claude Platform model ID as claude-opus-5-5. Confirm current API access and requirements in Anthropic's official documentation before changing a production integration.
How much does Claude Opus 5.5 cost?
Anthropic lists Opus 5.5 at $4 per million input tokens and $20 per million output tokens. Cache reads are $0.20 per million tokens and cache writes are $5 per million tokens. Fast mode is priced separately, and total task cost depends on usage, tools, retries, and workflow design.
Is Claude Opus 5.5 available now?
Anthropic says Claude Opus 5.5 is available across its platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure. Availability can depend on the selected provider, region, account, and deployment configuration.
Does Claude Opus 5.5 remove the need for human review?
No. Anthropic reports improvements in its safety evaluations but also describes limitations in measuring real-world behavior. Review model actions and outputs, use scoped tool permissions, and require approval for consequential or difficult-to-reverse actions.