Turn on extended thinking for hard debugging

The Claude 4 series (Opus, Sonnet, Haiku) supports interleaved thinking — the model generates a private reasoning trace between tool calls, not just before its final response. For multi-step debugging where you're handing the model several pieces of context in sequence (stack traces, logs, code reads), giving it a thinking budget is the cheapest win you'll find.

Via the Messages API:

response = client.messages.create(
    model="claude-opus-4-7",
    max_tokens=8000,
    thinking={"type": "enabled", "budget_tokens": 4000},
    messages=[...],
)

In Claude Code, the /think slash-command toggles thinking mode on for the current turn. For persistent use, set "thinkingMode": "interleaved" in settings.json.

Rule of thumb: thinking tokens cost the same as output tokens but produce no user-visible text. Budget enough for the model to actually reason (~2–8k for non-trivial debugging), not so much you waste money on trivial tasks.