Claude Haiku 5.5, Sonnet 5 and Sonnet 5.5 think by default, so the JSON eval
calls spend most of their output tokens on thinking (Haiku 5.5: 1643 out
tokens for a ~400 char answer). With the thinking budget at Off (the default)
send thinking={'type': 'disabled'}, or 'between_tools' on Sonnet 5.5 which
rejects 'disabled'. Claude 5.x has no thinking token budget any more, so
non-zero budgets keep the provider default. Opus 5 (leaks <thinking> tags
with thinking disabled), Opus 5.5 / Fable (thinking always on) and 4.x are
left alone.
Anthropic rejects unknown body fields, so client.completion() passes
'thinking' as a top-level litellm param instead of extra_body for Anthropic
models, and drops it on a BadRequest retry only when the error is about
thinking.
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>