What is Claude Opus 5.5?
Claude Opus 5.5 is Anthropic’s proprietary model for long-running agentic coding, computer use and knowledge work. Released September 22, 2026, it pairs a one-million-token context window with adaptive thinking, a 128,000-token maximum synchronous output and lower prices than Claude Opus 5.
Claude Opus 5.5 uses adaptive thinking with medium effort as the Claude API default. Higher effort can improve difficult work but also changes latency and cost, so teams should test the exact setting they intend to deploy.
QUICK VERDICT: Claude Opus 5.5 is a strong choice for demanding coding and professional workflows when Fable 5.1 is more capability or cost than the task needs. Its lower token price and improved efficiency are meaningful, but migrations must account for always-on thinking and changed tool-use behavior.
Field | Verified value |
|---|---|
Provider | Anthropic |
Release date | September 22, 2026 |
Availability | Active on the Claude API and supported cloud platforms |
License | Proprietary |
Model ID | claude-opus-5-5 |
Context window | 1 million tokens |
Maximum output | 128,000 tokens synchronously; 300,000 in Message Batches beta |
Modalities | Text and image input; text output |
Base API pricing | $4 input and $20 output per million tokens |
Best fit | Long-running agentic coding, computer use and knowledge work |
Last verified | September 24, 2026 |
What changed from Claude Opus 5
The most important changes are efficiency, communication and application behavior. Anthropic says typical workloads cost about 40% less to run than Opus 5 because Opus 5.5 combines lower token prices with lower token use. The company also reports output more than 30% faster at default settings.
Area | Claude Opus 5.5 change |
|---|---|
Price | $4 input and $20 output per million tokens, down from $5 and $25 for Opus 5 |
Cache reads | $0.20 per million tokens, down from $0.50 |
Reasoning | Adaptive thinking is always on; effort controls depth and defaults to medium |
Tool use | Forced tool use is not supported and returns an error |
Computer use | The older computer_20251124 tool version is not accepted on the Claude API or Google Cloud |
Communication | Anthropic reports clearer, more direct writing than Opus 5 |
Speed option | Fast mode offers up to 2.5 times the speed at $8 input and $40 output per million tokens |
These are not drop-in details. Applications that disable thinking, force a particular tool, reuse thinking blocks across models or stream text between tool calls may need code changes. Run the provider migration guide against the actual integration before changing a production model ID.
Claude Opus 5.5 capabilities
Anthropic positions Opus 5.5 for long and complex jobs rather than quick commodity generation. Its clearest strengths are codebase-wide migrations, audits, multi-step terminal work, computer use, research and professional analysis. The million-token window can hold large repositories or document sets, but context capacity is not proof that the model will retrieve every relevant detail reliably.
Capability | Practical implication |
|---|---|
Agentic coding | Useful for migrations, audits and multi-step repository work with tests and approval gates |
Knowledge work | Designed for research, financial analysis, reports and other multi-source professional tasks |
Computer use | Can operate interfaces through supported tools, but needs narrow permissions and recoverable actions |
Long context | Supports up to 1 million tokens; test recall and instruction retention at realistic lengths |
Adaptive thinking | The model chooses how much internal work to do within the configured effort level |
Large output | Supports 128K synchronous output and a 300K Message Batches beta for selected workloads |
A production agent is more than its base model. Prompts, retrieval, tools, permissions, memory, retries and verification determine whether the system succeeds. Evaluate the complete workflow, including recovery after a failed tool call and the amount of human review needed before accepting the result.
Benchmarks and evidence
Anthropic reports results in agentic coding, computer use and knowledge work, including 66.4% on Terminal-Bench 4.0 at xhigh effort, 81.8% partial on OSWorld 2.0 and 1846 Elo on GDPval-AA v2.1 at max effort. These are provider-reported evaluations with their own configurations and should be validated independently.
Pricing and access
Claude Opus 5.5 is available through the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS. Anthropic also makes the model available in its first-party Claude products, with access and usage limits depending on the plan.
Usage | Price per million tokens |
|---|---|
Input | $4 |
Output | $20 |
Five-minute cache write | $5 |
One-hour cache write | $8 |
Cache read | $0.20 |
Batch API | 50% discount on input and output |
Fast mode | $8 input and $40 output |
Token rates are only the starting point. Measure total accepted-task cost, including reasoning effort, cache behavior, long prompts, tool calls, retries, latency and human correction. A model with a higher list price can still be cheaper if it finishes in fewer steps, while an expensive reasoning setting can be wasteful on straightforward work.
Limitations and migration risks
- Adaptive thinking cannot be disabled, which changes latency and response structure compared with some earlier Claude integrations.
- Forced tool choice is not supported. Workflows that require one exact tool call must be redesigned and tested.
- Thinking blocks are tied to the model and conversation that produced them and should not be moved across models.
- Long-context capacity does not guarantee perfect recall, prioritization or citation accuracy across a million tokens.
- The strongest benchmark results often use high or maximum effort, which can differ from the API default in cost and speed.
- Computer use and long-running agents can make consequential mistakes, follow prompt injections or act outside an intended boundary.
- Vendor benchmarks and early-customer reports are useful evidence, but they are not a substitute for an independent test on your tasks.
High-stakes legal, financial, medical, security and scientific work needs qualified review. Restrict tools to the minimum required scope, require approval before irreversible actions, log the exact model and prompt version, and keep a rollback path.
How to evaluate Claude Opus 5.5
Build a frozen evaluation set from 20 to 50 real tasks. Include normal cases, edge cases, adversarial inputs and long-running work. Compare Opus 5.5 with Opus 5 and at least one neighboring model using equivalent tools, source material and acceptance criteria. Test both the default medium effort and the higher setting you would actually deploy.
Test area | Record |
|---|---|
Task completion | Pass or fail plus a written quality rubric |
Reliability | Repeated-run success rate and variance |
Tool use | Wrong calls, retries, recovery and permission-boundary failures |
Quality | Factuality, instruction adherence, citations and deliverable usefulness |
Efficiency | Wall time, input, output, cached tokens, tools and review time |
Migration | Code changes, silent output differences and regression failures |
Safety | Prompt-injection response, sensitive-data handling and irreversible actions |
Choose the least expensive configuration that clears the acceptance threshold with a safety margin. Re-run the suite whenever the model, system prompt, retrieval layer, tool definitions or approval policy changes.
Frequently asked questions
Is Claude Opus 5.5 the best AI model?
No model is best for every workload. Claude Opus 5.5 is worth testing for demanding agentic and professional work, but the right choice depends on the task, effort setting, cost, latency, modalities, deployment constraints and agent harness.
Is Claude Opus 5.5 open source?
No. Claude Opus 5.5 is proprietary and available through Anthropic and supported cloud platforms.
How much does Claude Opus 5.5 cost?
Base Claude API pricing is $4 per million input tokens and $20 per million output tokens. Cache reads cost $0.20 per million tokens, and Batch API input and output receive a 50% discount. Fast mode is priced separately.
What is Claude Opus 5.5 best used for?
Its strongest documented fit is long-running agentic coding, computer use and knowledge work. Start with a supervised pilot and compare total cost per accepted result, not just token prices.
Should I migrate from Claude Opus 5?
Only after testing the integration changes. Opus 5.5 costs less and Anthropic reports better speed and capability, but always-on thinking, tool-choice restrictions and response-shape differences can break assumptions in an existing application.
Related model guides
Compare with the Claude Opus 5 guide.
Compare with the Claude Fable 5.1 guide.
Browse the Best AI Models directory.
Official sources and update policy
Anthropic’s Claude Opus 5.5 announcement.
Claude Opus 5.5 model documentation.
Checked September 24, 2026. We update this guide when Anthropic changes the model specification, pricing, access, safety documentation or publishes material benchmark and deployment updates. Vendor benchmarks are attributed and are not presented as independent testing. No vendor payment or affiliate relationship determined inclusion.