Adaptive thinking and extended thinking in Claude

How Claude's reasoning modes work in 2026: adaptive thinking with the effort parameter on Opus 5 and Sonnet 5, versus extended thinking on Haiku 4.5.

Claude has two reasoning modes. Adaptive thinking - the current one - lets Claude decide whether and how deeply to reason before answering, with depth steered by an effort parameter that runs low, medium, high (the default), xhigh, max. Extended thinking is the older manual mode where you set a fixed budget_tokens. As of August 2026 only Haiku 4.5 still uses it.

Both modes cost the same way: thinking tokens are billed as output tokens. That is the fact that decides whether reasoning depth is a bargain or a waste on your workload.

#Adaptive thinking versus extended thinking

Extended thinking gave you one dial: a token budget for reasoning. You guessed how much a request needed and paid for that much whether it was warranted or not. Adaptive thinking moves that judgement to the model, with your effort setting as a guide rather than a hard allocation.

Two parameters are worth separating in your head. Anthropic's documentation puts it plainly: thinking controls whether Claude reasons in thinking blocks before answering; effort controls how much work Claude puts into the whole response, answer included.

You can also control what comes back. thinking.display is either "summarized", a readable summary of the reasoning, or "omitted", an empty field carrying an encrypted signature. On Fable 5, Opus 4.8 and Opus 4.7 the default is "omitted", and Fable 5 never returns its raw chain of thought. Hiding it does not make it free - you are still billed for the tokens.

#The effort ladder

Five rungs: low, medium, high, xhigh, max. The default is high on Opus 4.8 across all surfaces, and on Opus 5 and Sonnet 5 on the Claude API and in Claude Code. On Opus 5 the wire shape is an output_config object:

from anthropic import Anthropic

client = Anthropic()

msg = client.messages.create(
    model="claude-opus-5",
    max_tokens=8192,
    output_config={"effort": "max"},
    messages=[{"role": "user", "content": "Find the race condition in this scheduler."}],
)

print(msg.usage.input_tokens, msg.usage.output_tokens)

Note what is not in that request: no temperature, top_p or top_k. Those return a 400 on Opus 4.7 and later and on Sonnet 5. Effort is the depth control now; shaping behaviour is a prompting problem - see prompt engineering.

Practical guidance on the rungs, given the cost structure below:

• low - classification, extraction, routing, formatting, short factual lookups. Anything where the answer is short and the path to it is obvious.
• medium - summarisation, ordinary code generation, drafting. The rung most interactive products should sit on.
• high - the default. Multi-step analysis, debugging, agentic tool loops.
• xhigh and max - hard problems where a wrong answer is expensive and you can absorb the latency and the output bill. Verify the gain on your own evaluation set first.

#What thinking actually costs

Thinking tokens are billed as output tokens, count toward your max_tokens limit, and count toward your output-tokens-per-minute rate limit - even when display is "omitted". On the current models that is $10 per million on Sonnet 5, $25 on Opus 5 and $50 on Fable 5.

The arithmetic is stark for short-answer workloads, as the table below shows: on a request returning a 400-token answer, a few thousand tokens of reasoning can make the visible answer a tenth of what you paid for.

Those thinking-token counts are illustrations, not published averages - Anthropic does not publish typical reasoning lengths, and adaptive thinking varies them per request. Measure yours from the usage block on real traffic.

Two levers help. Prompt caching does nothing for thinking tokens - those are output - but it strips most of the input cost from the surrounding conversation, and the Batch API's flat 50% discount applies to thinking tokens like any other output. Both are worked through on the Claude API page, and cost per task across the four models is on the model comparison.

#The Opus 5 error you will hit

On Opus 5, thinking: {"type": "disabled"} is accepted only when effort is high or below. Combine it with xhigh or max and the API returns a 400. This is a deliberate breaking change from Opus 4.8, where the combination was allowed.

The logic is coherent once you see it: at the top two rungs the model's approach depends on reasoning in thinking blocks, so maximum effort with no thinking is a contradiction rather than a configuration. Pick one - drop to high, or leave thinking on.

#When deeper thinking helps, and when it burns money

Reasoning depth is not a quality slider. It is a trade of output tokens and latency for accuracy on a specific class of problem, and outside that class you are paying for nothing.

One migration trap: instructions written for older models often tell Claude to double-check itself - "include a final verification step", "use a subagent to verify". Anthropic advises removing those when moving to Opus 5, because they cause over-verification on a model that already reasons by default. You pay twice for the same caution. The same applies to prompts carried from Sonnet 4.6 to Sonnet 5.

The honest test is empirical. Run your evaluation set at low and at high, record accuracy, output tokens and p95 latency for both, and see whether the accuracy delta justifies the token delta. On most production workloads it does not. Patterns for varying effort per request type are covered under automation and Claude for developers.

#How Claude got here

Hybrid reasoning arrived with Claude 3.7 Sonnet in February 2025 - the first Claude model that could either answer immediately or think first, in one model rather than two. It shipped alongside the original Claude Code research preview, and its most-watched demonstration was a long-horizon run at playing Pokémon, which we cover in the Pokémon analysis. That model retired on 19 February 2026.

What followed was the control moving from the user to the model. Claude 4 and Sonnet 4.5 kept the manual budget; Opus 4.5 introduced effort in public beta in November 2025; Opus 4.6 and Sonnet 4.6 supported both modes and deprecated the older one; by Opus 4.7 and Sonnet 5 extended thinking was removed outright; Fable 5 made thinking mandatory. The full sequence is on the model index.

In the apps, extended thinking is available on every plan including the free tier. What you cannot do there is set an effort value directly - that remains an API and Claude Code control.

#Frequently asked questions