On Claude Opus 5, thinking is on by default. Omit the parameter and you get adaptive thinking — the model decides how much reasoning a request deserves. On Opus 4.8 and 4.7, omitting it meant no thinking at all. That reversal is the single most consequential default change in the release.
What adaptive thinking is
Older models took a fixed thinking budget in tokens. Adaptive thinking replaces that: the model decides, per request, how much to reason. There is no budget to tune — the depth is shaped by the effort level instead.
The old budget_tokens configuration is removed on Opus 5 and returns a 400 if you send it.
Why the default change bites
max_tokens is a single ceiling covering thinking and the visible response together. A request that ran without thinking on Opus 4.8, sized tightly around its expected answer, can now spend part of that budget on reasoning and truncate mid-sentence.
If you are migrating: raise max_tokens on every route that never set a thinking parameter, or set thinking: disabled explicitly to preserve the old behaviour.
Can you see the reasoning?
Raw thinking tokens are never returned on Opus 5. What you can get is a readable summary, by setting the thinking display option to summarized. The default is to omit it, which streams thinking blocks with empty text.
For a product that streams reasoning to users, the default looks like a long pause before anything appears — set the summary option explicitly if that is your interface.
If you disable thinking, know these two failure modes
Disabling thinking is permitted at effort high or below, but it introduces two behaviours worth planning for:
- Tool calls can arrive as plain text. The model occasionally writes a tool call into its visible response instead of emitting a structured call. The turn completes normally, no error is raised, and the tool never runs — a silent failure that is particularly nasty inside an agent loop.
- Internal tags can leak into the visible response.
The reliable fix for both is to turn thinking back on and use a lower effort level to control cost instead. If you must keep it off, allow the model a brief sentence before a tool call, and remove any instruction telling it not to reason — counter-intuitively, that instruction makes tag leakage worse rather than better.
More on Claude Opus 5
Start with the complete Claude Opus 5 guide for the overview, or go deeper:
Ready to go deeper?
Read the full Claude Opus 5 guideFrequently Asked Questions
Is extended thinking enabled by default on Claude Opus 5?
Yes. Omitting the thinking parameter runs adaptive thinking, which is a reversal from Opus 4.8 and 4.7 where omitting it meant no thinking. Since max tokens covers reasoning and response together, tightly-sized limits carried over from an older model can now truncate.
Can I turn thinking off on Claude Opus 5?
Yes, at effort `high` or below — pairing disabled thinking with `xhigh` or `max` returns a 400. Before you do, note that disabled thinking introduces two failure modes: tool calls occasionally appear as plain text and never execute, and internal tags can leak into output. Lowering the effort level is usually the better cost lever.
Can I see Claude Opus 5's reasoning?
Only as a summary. Raw thinking tokens are never returned on this model. Setting the thinking display option to `summarized` returns a readable summary; the default omits it, which streams empty thinking blocks and looks like a pause in a streaming interface.
