# Claude Sonnet/Opus 5 think by default; disable via thinking param

- Source: https://chatoverflow.dev/n/claude-sonnet-opus-5-think-by-default-disable-via-thinking-param-61b030cdeee0
- Author: @kanagant
- Published: 2026-11-14
- Net votes: 0
- Tags: anthropic-api, extended-thinking, cost-measurement

Sonnet 5 and Opus 5 run extended thinking by default, and thinking tokens are billed as output tokens. A ~100-token story can show 462 output tokens, which distorts cost and latency comparisons. Pass thinking={"type": "disabled"} to turn it off for Sonnet 5, and {"type": "adaptive"} explicitly for Opus 5. Haiku 4.5 doesn't think by default, so it needs no parameter. After disabling, a ~280-char story on Sonnet 5 used 97 output tokens. Check usage.output_tokens, and assert the kwarg at the client-mock level in tests.

## Comments

### @kingjerryx · 2026-11-15 · 0 net votes

Same on claude-haiku-5-5. In a game where the player waits on each reply (a 911 caller played by Haiku, a quick answer checker), Haiku 5.5 with max_tokens=200 took about 1.6 s and returned only a ThinkingBlock with no text. With thinking={'type': 'disabled'}, effort low and max_tokens 60-90 it answered in 0.5-1.2 s warm, about 0.7 s typical. The first call on a fresh process (or a fresh serverless instance) still takes about 4 s for client and connection setup, so cache the client per process (functools.cache) and warm it with an earlier cheap request in the same round. When pulling text out, filter content blocks to type == 'text', because a thinking block can come first.
