On Sonnet 5, turn thinking OFF for structured extraction

Here’s a default that surprised me and cost some clarity to untangle: on Claude Sonnet 5, omitting the thinking parameter runs adaptive thinking — it’s on by default. That’s the opposite of the Opus 4.x family, where omitting it means thinking is off.

For a chat or a hard reasoning task, adaptive thinking is exactly what you want. But for structured extraction — “read this page and emit these six fields as JSON” — there’s no reasoning to do, and adaptive thinking silently spends part of your max_tokens budget before the model emits the answer. On a small max_tokens, that can truncate the tool call you were forcing, and you get an empty or clipped result with no obvious cause.

The fix is explicit, and it’s Anthropic’s own documented starting point for extraction / classification work:

{
  "model": "claude-sonnet-5",
  "thinking": { "type": "disabled" },   // don't let adaptive thinking eat the budget
  "output_config": { "effort": "low" }, // cheapest, fastest — a separate knob
  "tool_choice": { "type": "tool", "name": "emit_fields" }
}

Two things worth internalizing:

• thinking and effort are separate knobs. Disabling thinking doesn’t make effort redundant — the pairing thinking: disabled + effort: low is the recommended config for extraction, not a belt-and-suspenders accident.
• A forced tool call composes with disabled thinking on the first-party API. The rule that forced tool_choice requires thinking: disabled is Bedrock-only; on api.anthropic.com it just works either way.

Net: for extraction, disable thinking on purpose. Don’t inherit a default that was tuned for a different job.