Claude Opus 5

A New Opus Generation Built Around Sustained Work

Anthropic described Opus 5 as a step-change rather than an incremental update. The largest gains over Opus 4.8 were in deep reasoning, agentic and long-horizon tasks, and the ability to convert additional test-time compute into better results.

That positioning matters because Opus 5 was not designed only for single-turn question answering. It was built for workflows where the model must remain useful across a sequence of decisions: inspect a codebase, delegate to subagents, call tools, verify changes, work through long documents, or maintain a complex enterprise task over an extended session.

The model has a 1,000,000-token context window and supports up to 128,000 output tokens. It accepts text and images and generates text. Thinking is on by default, with high as the default effort level.

Opus 5 has since been superseded by Opus 5.5 as Anthropic’s recommended general frontier model. That makes Opus 5 especially useful today as an operational baseline for teams with existing prompts, evaluations, tool loops, and production behavior built around the first Opus 5 release.

• Model ID: claude-opus-5.
• Released July 24, 2026.
• 1M-token context window and 128K standard output.
• Text and image input with text output.
• Thinking on by default with configurable effort.
• Legacy model with Opus 5.5 as the recommended successor.

Frontier Capability Is Most Valuable When Several Skills Interact

Enterprise workflows rarely ask for only one capability. A financial analysis may require reading long documents, interpreting tables, checking calculations, maintaining a set of constraints, and producing a polished artifact. A frontend reconstruction task may require visual inspection, code generation, iteration, and comparison against the source image.

Anthropic specifically highlights Opus 5 improvements in vision, office and document tasks, and long-context instruction following. The model can analyze charts, documents, diagrams, and UI visuals and can work across complex multi-sheet spreadsheets and structured slide decks.

The practical evaluation question is not whether the model can describe the input. It is whether it produces a result that survives downstream verification: formulas calculate correctly, slides preserve the intended argument, visual recreations match the reference, and long documents remain internally consistent.

Large Context Supports Repository and Enterprise State

A million-token window gives an agent room for much more than a long prompt. It can hold source code, documentation, conversation history, image inputs, tool results, plans, research evidence, enterprise policies, and intermediate work in one active context.

Anthropic emphasizes consistent instruction following, tool calling, and reasoning throughout the long context. That matters because raw capacity is useful only when the model can still identify relevant instructions and evidence deep in the window.

For batch workloads, Anthropic also exposes a beta maximum output of 300,000 tokens using the Message Batches API. Standard interactive output remains 128,000 tokens.

Large context should still be managed deliberately. Prompt caching, retrieval, compaction, and context editing can reduce cost and noise even when the model technically has enough room for everything.

Change Tools Without Rebuilding the Whole Session

Anthropic introduced mid-conversation tool changes in beta for Opus 5. This allows an application to add or remove tools between conversation turns while preserving the prompt cache rather than treating the tool list as fixed for the entire session.

That is useful for agents whose permissions and available actions evolve during a task. A coding agent might gain a deployment tool only after tests pass. A research workflow might add a specialized data source only when the investigation reaches a specific stage.

Opus 5 also supports a research-preview fast mode on the Claude API. It provides lower-latency inference at separate premium pricing, creating another evaluation dimension: standard Opus 5 quality and economics versus the same model family in a latency-optimized mode.

Other relevant capabilities include text and image input, tool use, prompt caching, large-context processing, code-oriented workflows, and batch generation with the higher-output beta path.