How to Prompt Claude Opus 5 (And Why You Shouldn't Have To)

Claude Opus 5 just launched. Its own documentation says thinking is on by default, effort levels matter more than ever, and old prompts need tuning. This post covers what actually changed and what you need to do about it.

What changed in Claude Opus 5

Anthropic changed several default behaviors in Opus 5. The biggest shift is that thinking is on by default. On Opus 4.8, you had to opt in with thinking: {"type": "adaptive"}. Now it runs automatically. That single change affects everything from token budgets to response structure.

Here is what else changed:

• Effort levels matter more. You get 5 levels from low to max. Opus 5 converts additional effort into better results more reliably than any earlier model.
• Breaking change with thinking disabled. Disabling thinking at xhigh or max effort now returns a 400 error. This was allowed on Opus 4.8.
• Mid-conversation tool changes (beta). You can add or remove tools between turns without breaking the prompt cache.
• Lower prompt cache minimum. Now 512 tokens, down from 1,024 on Opus 4.8.
• Fast mode. Available for Opus 5 on the Claude API only.
• Context window. 1M tokens, both the default and the maximum.
• Pricing unchanged. $5 per million input tokens, $25 per million output tokens.

The full list is in Anthropic's official Opus 5 changelog.

What the prompting guide actually says

Anthropic also published a dedicated prompting guide for Opus 5, and it flags real behavioral differences that will affect your results if you ignore them.

The biggest one: default responses run longer than prior Opus models. The effort parameter controls thinking depth, not response length. If you want shorter answers, you need to say so explicitly. Something like "Keep responses focused, brief, and concise" works. Without that instruction, Opus 5 will write longer reports, longer code files, and narrate its agentic work step by step.

The second pattern that catches people off guard is over-verification. Opus 5 already verifies its own work internally. If your existing prompts include instructions like "include a final verification step" or "double-check your answer," those now compound with the model's built-in self-correction. The result is wasted tokens and looping behavior. Remove those instructions.

Here are the other behavioral changes worth knowing:

• Aggressive subagent delegation. Opus 5 spawns subagents more readily than older models. Cap delegation explicitly if cost matters.
• Narration during agentic work. It announces what it is about to do before doing it. You can tune this down with explicit guidance on communication cadence.
• Thinking disabled artifacts. With thinking turned off (only allowed at effort high or below), tool calls can leak into text output and internal XML tags can appear. The recommendation is to keep thinking enabled and use lower effort levels instead.

The documentation treadmill

This is one model from one lab. Anthropic alone shipped Opus 4.8 and now Opus 5 with different behavioral defaults. Multiply by OpenAI, Google, and every other lab, and you have a permanent homework assignment just to keep your prompts current.

We covered the same pattern when Fable 5 launched. Anthropic's own fix for Fable 5's cost was a specific delegation pattern that users had to learn manually.

How Kosmo handles this for you

Kosmo's knowledge vault already has the Opus 5 documentation and prompting guide ingested. It was added within hours of the release.

When you compile a prompt targeting Claude, Kosmo applies the correct patterns automatically. It applies the right effort guidance, the right verbosity constraints, and removes leftover verification scaffolding from older models. You describe what you want in plain language, pick your target tool, and Kosmo compiles a prompt that follows the current documentation for that model. No manual tuning, no reading migration guides. The same way Kosmo already handled the Fable 5 delegation pattern, it now handles Opus 5's behavioral changes. Ready to stop fighting documentation? Try Kosmo today.

Frequently asked questions

Do I need to rewrite all my Claude prompts for Opus 5?Not from scratch. But you should remove any explicit verification instructions like "verify your work" or "double-check". You should also add a conciseness instruction if responses are running too long. The biggest change is that thinking is on by default, which means max_tokens now covers both thinking and response text. Review that limit.

What effort level should I use? Start at the default, high. Step down to low or medium where quality holds to save tokens and latency. Step up to xhighfor demanding coding and agentic work. Anthropic's own guide says to run an effort sweep on your evals rather than guessing.

How does Kosmo stay current when models change this often? Kosmo's knowledge vault is updated with new documentation as labs release it. When you compile, Kosmo pulls the current prompting guide for your target model. You do not need to track releases yourself.