Migrating to Claude Opus 5 is mostly a one-line change followed by an afternoon of prompt re-tuning. The code changes are few; the prompt changes are where the value is, and several of them are deletions rather than additions.
Things that will break
1. Requests that never set a thinking parameter. Thinking is now on by default, and your max-token budget covers thinking plus response together. Tightly-sized limits truncate. Raise max_tokens, or set thinking to disabled explicitly.
2. Disabled thinking above high effort. Combining disabled thinking with xhigh or max returns a 400. The check runs on every request, so audit every call site rather than just the first.
3. Anything still carrying removed parameters. temperature, top_p, top_k, the old budget_tokens thinking config, and assistant-turn prefills all return 400. These were already removed before Opus 5 — but if you are jumping from Opus 4.6 or earlier, this is where you find out.
Things to delete from your prompts
- Verification instructions. "Double-check your work", "re-verify before responding", and separate verification passes in your harness. Opus 5 verifies natively; these now cause redundant effort. This inverts the usual prompting advice — it is a carve-out specific to this model.
- "Delegate more" guidance written for Opus 4.8, which under-reached for subagents. Opus 5 over-reaches. Replace it with an explicit ceiling.
- Forced progress scaffolding such as "summarise after every three tool calls". Opus 5 narrates on its own.
Things to add
- A conciseness instruction. Default responses run long, and effort does not shorten them.
- A deliverable-length instruction if the model writes files — reports and documentation trend toward padding.
- A scope-discipline instruction if you see it widening the task beyond what you asked for.
- A subagent cap if your harness supports delegation.
- Refusal handling. Check
stop_reasonbefore reading content, and consider enabling server-side fallback so declined requests re-run on Opus 4.8 automatically.
Things to re-measure
- Effort per route. Sweep
medium,high, andxhigh— inherited defaults are rarely right. - Cacheability. The minimum drops to 512 tokens, so re-check prompts you had written off.
- Rate limits. Opus 5 has its own bucket and no Priority Tier coverage.
More on Claude Opus 5
Start with the complete Claude Opus 5 guide for the overview, or go deeper:
Ready to go deeper?
Read the full Claude Opus 5 guideFrequently Asked Questions
How hard is it to migrate from Opus 4.8 to Opus 5?
The code change is usually just the model ID string. The real work is prompt re-tuning: raising max tokens where thinking is now on by default, deleting verification instructions, capping subagent delegation, and adding a conciseness instruction. Budget an afternoon, not a sprint.
What is the most common migration mistake?
Keeping instructions that tell the model to verify or double-check its work. Opus 5 does that natively, so those instructions now produce redundant effort — more tokens and more latency for no quality gain. Removing them is the highest-value single change.
Do I need to change my max tokens setting?
On any route that previously omitted the thinking parameter, yes. Thinking is on by default on Opus 5 and shares the max-token budget with the visible response, so a limit sized tightly around the expected answer on Opus 4.8 can now truncate mid-response.
