Claude Opus 5

Overview

Anthropic introduced Claude Opus 5 on July 24, 2026. Built for complex agentic coding and enterprise work, it provides greatly improved performance for the same cost as its predecessor, Opus 4.8. The model ID is claude-opus-5 and pricing is $5 per million input tokens and $25 per million output tokens. It is the new default model on Claude Max and the strongest model on Claude Pro.

Key features

• Verifying its work

  On long agentic runs, users often had to build verification steps into their prompts because the model would not check its own results. Claude Opus 5 is much stronger at verifying its work and iterating carefully until it succeeds. The official migration guide recommends removing existing self-check scaffolding, since explicit verification instructions now cause over-verification.

• Thinking on by default

  On Claude Opus 4.8, requests without a thinking field run without thinking. On Claude Opus 5, the same requests run with adaptive thinking enabled by default. Thinking tokens count against max_tokens, so that limit needs revisiting; to preserve the old behavior, set thinking: {"type": "disabled"} explicitly.

• Full effort range

  Claude Opus 5 supports the full effort range — low, medium, high, xhigh, and max — which customers can use to optimize for intelligence or conserve tokens for faster and cheaper results. The default is high on the Claude API and Claude Code, and max is worth testing for capability-critical work.

• Mid-conversation tool changes (beta)

  Changing the tool list partway through a conversation used to invalidate the prompt cache, adding cost and latency. With the mid-conversation-tool-changes-2026-07-01 beta header on the Claude Platform, developers can now change which tools Claude can use without invalidating the prompt cache.

• Automatic fallbacks (beta)

  Requests flagged by the safety classifiers previously had to be handled by the application. With the server-side-fallback-2026-07-01 beta header and fallbacks: "default", users can now choose to have requests that are flagged on Opus 5 (or Fable 5) automatically route to another model. In Claude.ai, Claude Code, and Claude Cowork, any flagged requests will fall back to Opus 4.8 by default.

• Lower prompt caching minimum

  The minimum cacheable prompt length is now 512 tokens, down from 1,024, so shorter prompts can create cache entries with no code changes.

Benchmarks

Pricing & availability

Notes

• Existing Opus 4.8 code does not work unchanged — requests without a thinking field now run with thinking, and thinking: {"type": "disabled"} combined with xhigh or max effort returns a 400 error. To keep thinking disabled, lower effort to high or below.
• Response length and instruction following changed — Claude Opus 5 calibrates response length to task complexity, so default responses run longer, and it follows instructions more literally with less implicit generalization. Prompt explicitly for conciseness or a target length.
• Spawns subagents more readily — it delegates to subagents more readily than Opus 4.8, so give explicit guidance or cap the subagent count where that matters.
• Built-in progress updates — agentic traces include progress updates, so scaffolding that forces interim status messages can be removed.
• Safeguards — Claude Opus 5’s safeguards are designed to allow beneficial uses of the model in both cybersecurity and biology. They allow Opus 5 to find vulnerabilities in source code, but block “binary-based” vulnerability scanning; the Cyber Verification Program facilitates cybersecurity work that would otherwise be impeded by the model’s safeguards. Opus 5 remains behind Mythos 5 on cybersecurity exploitation tasks and autonomous biology research.
• Opus 4.1 retirement — claude-opus-4-1-20250805 is deprecated and will be retired on August 5, 2026; migrate to Claude Opus 5 before then.