Claude Sonnet 5 migration checklist

Anthropic positions Claude Sonnet 5 as an upgrade path from Sonnet 4.6, but its docs call out changed sampling support, adaptive thinking behavior, larger context/output limits, and tokenizer differences. Why it matters: Model swaps can quietly change cost, latency, and prompt behavior even when the API surface looks familiar.

Try this: Before routing production-like agent tasks to Sonnet 5, remove unsupported sampling parameters and rerun a small token, latency, and output-quality comparison against the current model.