Extended Thinking shipping

Slow Claude down on hard things.

On hard problems — multi-step reasoning, tricky math, gnarly code — extended thinking gives Claude an explicit budget of thinking tokens before it produces the user-visible answer. The thinking trace is returned alongside the answer so you can inspect or hide it. Trades latency and tokens for quality.

The decision rule.

Pass a budget.

msg = client.messages.create(
    model="claude-opus-4-7",
    max_tokens=4096,
    thinking={"type":"enabled","budget_tokens":2000},
    messages=[{"role":"user","content": HARD_QUESTION}],
)
# msg.content has both `thinking` blocks and `text` blocks.
# `thinking` is for inspection; show only `text` to end users.

The model decides how much of the budget to actually use. A 2,000-token budget is not a guarantee of 2,000 thinking tokens — just an upper bound.

Show or hide the trace.

• Hide it in production for end users — they want answers, not transcripts.
• Show it in developer-facing tools and "explain your reasoning" experiences.
• Stream it so users see something happening while the model thinks.
• Log it for evals and debugging.

Combine.

• Tool Use — Claude can think before deciding which tool to call.
• Agent SDK — agents on hard goals benefit from extended thinking on the planning step.
• Prompt Caching — thinking does not invalidate cached prefixes; the system prompt and docs stay cached.

Related products.

Every Anthropic surface has its own page on this guide.