Adaptive thinking: how Claude decides how hard to think
In brief
Claude doesn't apply the same effort to every question. Here's what adaptive thinking is, how it works, and why it matters for the outputs you get.
Contents
When you ask Claude a simple question — "What's the capital of France?" — it answers immediately. When you ask it something complex — "What are the second-order effects of raising our enterprise pricing by 20%?" — it should take more care. And it does, automatically.
That's adaptive thinking: Claude's ability to calibrate how much reasoning effort to apply based on what the question actually requires.
How it works
Claude assesses the complexity and ambiguity of each request and adjusts its internal process accordingly. Simple, factual questions get direct answers. Nuanced, multi-part, or high-stakes questions get more deliberate internal processing before a response is generated.
On Claude Opus 5, Claude Sonnet 5 and Claude Fable 5 this is the default and requires no configuration. On Claude Opus 4.8, 4.7, 4.6 and Sonnet 4.6 it is available but off until you set thinking: {type: "adaptive"}.
Adaptive thinking replaced extended thinking, the older arrangement where you explicitly enabled reasoning and handed the model a fixed token budget. That configuration — type: "enabled" with budget_tokens — is not accepted on the current models at all. You no longer decide whether Claude thinks. You influence how much it spends, using the effort parameter.
Why it matters in practice
Two things to understand:
The quality floor is higher than it used to be. Early language models applied roughly the same process to everything, which meant complex questions often got shallow treatment. Adaptive thinking means Claude is more likely to give appropriately careful answers to questions that need them — without you having to explicitly ask for more effort.
Complex questions still benefit from explicit framing. Adaptive thinking doesn't mean Claude always knows what a question requires. A question like "Should we expand into the European market?" looks like a business strategy question but might need very different treatment depending on your company's stage, resources, and risk tolerance. Giving Claude more context — "We're a 30-person SaaS company, €2M ARR, currently UK-only" — helps it calibrate more accurately.
Think of adaptive thinking as the default setting that gets you most of the way there. Clear context and good prompting take you the rest of the way.
One control worth knowing. effort — low, medium, high (the default), xhigh, max — steers how much Claude spends across the whole response, thinking included. Raise it for long agentic and coding work; lower it for high-volume routine tasks where the deliberation is not returning anything. See when to raise and lower effort for the workload-by-workload version.
Note also that thinking cannot always be switched off: Claude Opus 5 rejects thinking: {type: "disabled"} at xhigh and max effort, and Claude Fable 5 rejects it entirely.
Further reading
- Adaptive thinking documentation — how adaptive thinking works in the API
- Harnessing Claude's intelligence — how to get the most from Claude's reasoning capabilities
- Effort — the parameter that steers how deeply Claude thinks on current models
- Extended thinking documentation — the superseded manual configuration, still used by older models