Claude Opus 5

What is Claude Opus 5?

Claude Opus 5 is a general-purpose multimodal language model from Anthropic. It is designed for work where the model must reason through difficult problems, maintain a large amount of context, write or review code, and complete several steps with tools or external instructions. Its canonical Claude API model ID is claude-opus-5.

The model is multimodal in its input handling: it accepts text and images, including supported visual document content, but its direct output is text. It does not natively generate images, audio, or video. This makes it suitable for interpreting diagrams, screenshots, image-heavy documents, and visual PDF content while keeping the response in text or code.

Anthropic positions Opus 5 for complex agentic coding, enterprise work, deep reasoning, and long-running workflows. “Agentic” here refers to applications in which the model plans or performs multiple steps, often calling tools, inspecting results, revising its approach, and continuing until a task is complete.

Where Claude Opus 5 fits in Anthropic’s lineup

Claude Opus 5 is the high-capability Opus-tier model in Anthropic’s Claude model family. Its capabilities and price place it above lower-cost options intended for simpler or higher-volume generation. The trade-off is that Opus 5 is better suited to difficult, extended tasks than to routine requests where low latency and low token cost matter more.

Its current status is important for anyone choosing it for a new application. Anthropic classifies Claude Opus 5 as an active legacy model and recommends migration to Claude Opus 5.5. The published retirement commitment is not sooner than July 24, 2027. “Active legacy” means the model remains usable, but it should not automatically be treated as Anthropic’s preferred long-term choice for new deployments.

Context window and output limits

Claude Opus 5 has a 1-million-token context window. The context window is the total amount of information the model can consider for a request and its surrounding conversation, including instructions, documents, tool results, and generated content. In practical terms, this makes the model suitable for large repositories, lengthy contracts, extensive research collections, and long-running coding sessions.

The standard maximum output is 128,000 tokens. That is a maximum rather than a required response size: ordinary answers may be much shorter. Anthropic’s Message Batches API can support up to 300,000 output tokens in beta when the required output-size header is used. This higher batch limit should not be confused with the standard synchronous output limit.

Anthropic lists May 2026 as the model’s reliable knowledge cutoff and training-data cutoff. Web search and other retrieval tools can provide newer information when available, but retrieved information is separate from the model’s built-in knowledge.

Reasoning, coding, and tool capabilities

Claude Opus 5 uses adaptive thinking. Its default effort level is reported as high, with additional effort settings available to balance response quality, latency, and token use. Higher reasoning effort can be useful for complex debugging, planning, legal or financial analysis, and tasks where the model must compare several possible approaches. Lower effort may be more appropriate when speed and cost are more important.

The model supports tool and function use, allowing an application to give it structured tools and receive requests to call them. This can support workflows such as searching a knowledge base, reading files, running an application-specific operation, or coordinating multiple steps. Tool use does not mean that every deployment automatically grants access to external systems; the application or platform must provide and authorize the tools.

Claude Opus 5 also supports structured outputs. These allow applications to request responses that follow a defined structure rather than relying entirely on free-form prose. Structured responses are useful for extracting fields from documents, returning workflow decisions, or passing model results to software. The model supports streaming responses, prompt caching, and the Message Batches API as well.

For software engineering, the combination of long context, adaptive reasoning, tool use, and text generation is particularly relevant. It can be used for code generation, code review, repository-level analysis, debugging plans, documentation, and multi-step engineering agents. The supplied research supports these use cases, but it does not provide independent benchmark results, so performance should be evaluated against a representative workload rather than inferred from the model tier alone.

Pricing, caching, and cost trade-offs

Standard Claude Opus 5 pricing is $5 per million input tokens and $25 per million output tokens. Input tokens include the material sent to the model, while output tokens are generated responses. Long prompts, large documents, extensive tool results, and very detailed responses can therefore affect costs in different ways.

Prompt caching provides separate rates for reusable input context. Five-minute cache writes cost $6.25 per million tokens, one-hour cache writes cost $10 per million tokens, and cache reads cost $0.50 per million tokens. Caching can be useful when an application repeatedly sends the same large instructions, codebase context, or reference material. The benefit depends on whether the cached context is reused enough to offset the write cost.

Batch API requests receive a 50% discount on input and output pricing. Batch processing is more appropriate for workloads that do not require an immediate interactive response, such as large-scale document classification, offline analysis, or queued code and content processing.

These prices make Opus 5 a poor fit for routine, high-volume generation when a less capable and less expensive model can meet the quality requirement. Its cost is easier to justify when avoiding errors, preserving a very large context, or completing a complicated multi-step task has greater value than minimizing per-request spend.

Supported inputs and outputs

Web search is supported on the Claude API, Claude Platform on AWS, and Microsoft Foundry, but not on Amazon Bedrock according to the supplied deployment notes. Availability also depends on organization settings and the specific deployment surface. This distinction matters when the same model is accessed through different cloud providers.

Best use cases for Claude Opus 5

• Complex agentic coding: analyzing a large codebase, planning changes, reviewing implementation details, and coordinating several tool calls.
• Long-context research: comparing many documents or maintaining continuity across a long investigation.
• Enterprise analysis: working with detailed operational, legal, financial, or technical material where reasoning quality matters.
• Document production: drafting, revising, and structuring substantial documents while preserving requirements from a large source set.
• Code review and debugging: examining broad project context rather than isolated snippets.
• Tool-using applications: building workflows that require structured outputs, external tools, caching, and multiple stages.

The model’s large context does not remove the need for good application design. Irrelevant documents, poorly controlled tool permissions, and unverified retrieved information can still reduce reliability. For high-stakes uses, outputs should be checked against authoritative sources and tested against realistic examples.

Limitations and when to choose another model

Claude Opus 5 is not intended for native image, video, audio, speech, transcription, or embedding generation. An application requiring those outputs will need a different specialized model or a separate media service. Fine-tuning is also not listed as a supported capability for this model.

Its token pricing and high reasoning orientation make it less suitable for simple classification, short routine answers, or very large volumes of low-risk text generation. A faster, lower-cost model may be the better choice when the task has a small context, predictable structure, and limited reasoning requirements.

Opus 5 can also be slower or more expensive when configured for higher reasoning effort. That trade-off is useful for difficult tasks but unnecessary for every request. Applications can reserve it for escalated cases while routing straightforward work to another option, provided the quality and context requirements are still met.

Finally, its legacy status should influence procurement and engineering decisions. Existing users may continue using it while it remains available, but teams beginning a new project should evaluate Anthropic’s recommended Claude Opus 5.5 as well. The relevant comparison is not simply which model has the higher label; it is whether the successor provides the required quality, context, tools, and compatibility at an acceptable cost.

Bottom line

Claude Opus 5 is best understood as a high-end text-output model for difficult, context-heavy work. Its defining practical advantages are the 1-million-token context window, adaptive thinking, strong coding orientation, image understanding, tool support, structured outputs, caching, and batch processing. Its main disadvantages are cost, potentially higher latency at greater reasoning effort, lack of native media generation, and its current active-legacy status.

Choose it when a task benefits from sustained reasoning across large inputs or several tool-driven steps. Choose a less expensive or more specialized option when speed, volume, media generation, or long-term model availability is more important than maximum general-purpose reasoning capacity.