Swapping the model name
is not the migration.
Sonnet 5 launched on 30 June 2026, Opus 5 on 24 July. Our pages on what Opus 5 changed and Sonnet 5’s real pricing cover what the models are; this is the other half: the checklist for actually moving a team onto them without a confusing three-weeks-later production surprise.
The 5 series is materially more model for the money: million-token windows, 128,000-token maximum output, thinking managed adaptively. It also ships two genuine breaking changes and several moved defaults, which is why the migration deserves a checklist rather than a model-string swap.
Re-count everything: the tokenizer changed
Sonnet 5’s new tokenizer produces roughly 30% more tokens for the same text. Hard-coded token budgets, context calculations and cost estimates built on the old tokenizer are wrong after migration. Re-test with real outputs on your actual content rather than applying the 30% as a flat correction; it varies by content type.
Remove manual sampling parameters before they remove themselves
Sonnet 5 returns a 400 error on manual sampling parameter requests. Pipelines setting temperature or top-p from older models break, loudly, on day one. Audit for these calls first; it is the cheapest item on the list and the most certain to bite.
Review effort defaults, or inherit the wrong ones
Sonnet 5 switches on adaptive thinking by default; Opus 5 has thinking on by default and only allows disabling it at effort high or below. Teams coming from models where thinking was off can see cost and latency rise from the upgrade alone. Set effort per workflow as part of the migration; the effort guide covers how.
Check the output ceiling and the leftovers
Both models support 128,000-token maximum output; requests written for smaller max_tokens values deserve a review rather than a copy-forward. And note what was retired around the launches: fast mode for Opus 4.7 was removed the day Opus 5 shipped. Migrations are also the moment to stop depending on things scheduled to disappear.
Use the pricing window to load-test
Sonnet 5’s introductory $2/$10 per million tokens runs until 31 August 2026. That window is the cheap period in which to run your real workloads against the new model and let the new tokenizer arithmetic settle into your budget figures. After the window, you will be doing the same testing at full price.
Audit sampling parameters, re-test token budgets, set effort defaults, review max_tokens, then load-test before 31 August. Each step is small; skipping one is how migrations surface as confusing production issues three weeks later.
Things people
usually ask.
You can, but stagger them if capacity is tight, and start with Sonnet 5: its introductory pricing window ends 31 August, so its load-testing is the time-sensitive half. The checklist is the same for both.
Manual sampling parameters. Sonnet 5 rejects them with a 400 error, so any pipeline setting temperature or top-p fails immediately. The tokenizer change is subtler: nothing errors, but budgets and costs drift about 30% until recalculated.
Mostly, but prompts that assumed a thinking toggle or a particular effort behaviour may run slower, dearer, or with different depth until effort levels are set deliberately. Treat prompt review as part of the migration, not an afterthought.