Claude Sonnet 5 pricing, plans and limits

Explicitly set thinking mode

Pass thinking: { type: 'disabled' } to turn off adaptive thinking, or leave it default for adaptive reasoning.

Adaptive thinking runs by default on every request, unlike Sonnet 4.6, and consumes part of the max_tokens budget.

State an explicit reporting bar for code review

"Report every bug you find, including low-severity ones, even if below your usual bar."

Sonnet 5 follows instructions more literally and may silently drop low-severity findings it judges not worth reporting unless told otherwise.

Adaptive thinking on by default

Requests without an explicit thinking field ran without thinking on Sonnet 4.6 but run with adaptive thinking on Sonnet 5, consuming part of the max_tokens budget for the same request.

Workaround: Pass thinking: { type: 'disabled' } explicitly for requests that need the old no-thinking behavior, and re-check max_tokens limits for workloads migrated from Sonnet 4.6.

Sampling params now error at non-default values

Setting temperature, top_p, or top_k to anything other than the default returns a 400 error, unlike on Sonnet 4.6 where these were accepted.

Workaround: Remove these parameters from request payloads during migration, or explicitly pass the default value.

New tokenizer inflates token counts ~30%

The same input text produces roughly 30% more tokens than under Sonnet 4.6's tokenizer, so usage fields and pre-flight token counts are not directly comparable across generations.

Workaround: Recount tokens against Sonnet 5 directly rather than reusing cached counts from Sonnet 4.6, and rebudget max_tokens accordingly.

Fewer reported low-severity code review findings

The model investigates code as thoroughly as before and finds the same bugs, but converts fewer investigations into reported findings when it judges them below the stated severity bar, which can look like reduced recall even though detection ability improved.

Workaround: Explicitly instruct the model to report every finding regardless of severity when running code review or audit tasks.