Just went through the Anthropic migration guide for Opus 4.7 and there are more gotchas than the announcement implied. Curious if others have run into these.

Three things that break on a blind model ID swap:

1. **Old extended thinking payloads** — `{type: “enabled”, budget_tokens: N}` returns `400`. New format is `{type: “adaptive”, effort: “high”}`

2. **Custom sampling params** — non-default `temperature`, `top_p`, `top_k` all return `400` now

3. **Thinking text hidden by default** — reasoning still happens but you have to opt in with `display: “summarized”` to see it

Also worth noting: the updated tokenizer can use up to 1.35x more tokens for the same input. So “same price” doesn’t necessarily mean same bill.

The capability improvements look real for coding/agent workloads. But this feels like a migration that deserved a deprecation window, not a same-day hard cutover.

Anyone found a clean way to handle the sampling parameter removal? Especially for use cases that needed deterministic-ish output via `temperature=0`?

*Reference: Claude API migration guide (April 2026)*