Claude Opus 5: code switch alone breaks on 7 points

The article frames this as a version-history problem rather than a single bad release. Sampling parameters, a fixed thinking budget, and assistant prefill were all accepted in earlier Claude models, and the guide now marks them as removed or discouraged. Correct fixes have also flipped back and forth across 4.7, 4.8, and Opus 5, which is why the author suggests starting by suspecting the sentences you added for the previous model rather than the model itself.

The behaviors that don't raise an error are subtler. Opus 5 keeps responses long even at lower effort, verifies its own answers, and delegates to subagents more often, so leftover instructions written for 4.8 — asking for more delegation or for final verification — can push it into over-verification. Sonnet 5 reads instructions literally, which makes carried-over style notes like "be concise" bite harder, and a "report only critical issues" instruction can lower code-review recall. The author also notes that moving to the top-tier Claude Fable 5.1 adds three more constraints around disabled thinking, tool_choice, and edited conversation history.

Taken together, the practical stakes for teams already running Claude in production lie in prompt hygiene and measurement discipline. Whether a migration goes smoothly appears to hinge on auditing inherited prompts and re-checking token counts with count_tokens before trusting old budget assumptions, since the pricing change and the tokenizer change pull costs in opposite directions.