What Is Reasoning Effort in AI, and What Does It Cost?

Reasoning effort is a setting that tells a model how much internal thinking to do before it answers. Set it low and the model responds faster and cheaper with less deliberation. Set it high and it works through the problem more thoroughly, spends more tokens doing it, and bills you for them. It is not a quality knob in the sense of "better writing", it is a depth knob, and on most workloads the difference between the cheapest and most expensive setting is a multiple, not a rounding error.

Every major provider now exposes some version of this control. None of them call it the same thing.

The vocabulary problem

Here is where the dial lives on each of the three APIs most people building with AI actually touch, taken from vendor documentation rather than secondary write-ups.

Three different parameter paths, three different value sets, three different defaults. Anthropic's docs are explicit that effort: "high" "matches the API default", while OpenAI's guide states that "gpt-5.5 defaults to medium reasoning effort". A port from one provider to another that silently keeps "the default" is changing the depth of reasoning, not just the model.

That is the first practical thing to know: there is no neutral setting. Omitting the parameter picks one for you, and the one it picks differs by vendor.

What actually changes between levels

OpenAI's documentation puts it plainly: "Lower effort favors speed and lower token usage, while at higher effort the model thinks more completely to provide higher quality responses."

Underneath that sentence are three real effects.

Token spend. Reasoning tokens are billed as output tokens. They also occupy space in the context window. A high-effort answer to a short question can produce far more billed output than the visible response suggests, which is why a bill can climb while your logged answer lengths stay flat.

Latency. Thinking happens before the first visible token. Higher effort means a longer wait before anything appears, which matters enormously for chat interfaces and not at all for a nightly batch job.

Behaviour, not just depth. On Anthropic's adaptive-thinking models the docs note that at lower effort the model "may skip thinking entirely on easy inputs". That is a meaningful difference from a fixed token budget, where the model thinks on every request whether or not the request warrants it.

How this differs from a reasoning model

Worth separating two ideas that get blurred. A reasoning model is a model trained to produce an internal chain of work before answering. Reasoning effort is the control surface that decides how much of that work happens on a given request.

You can have a reasoning model with the dial turned to minimum, in which case it behaves closer to a conventional model. You cannot turn the dial up on a model that was never trained to think. The model is the capability, the effort setting is the throttle.

Measuring what a level costs you

Do not guess. Both major providers expose the number directly.

On OpenAI, the reasoning token count appears in the response usage object under output_tokens_details. On Anthropic, the docs point at usage.output_tokens_details.thinking_tokens, which "reports how many of the billed output tokens were internal reasoning", with the caveat that when streaming, the breakdown only arrives on the final message_delta event.

A ten-minute experiment worth running before you pick a default:

For a representative sample of 50 real requests:
  for each effort level in [low, medium/high, max]:
    record: reasoning tokens, total output tokens,
            time to first token, pass/fail against your rubric
Then plot pass rate against cost per request.

The shape of that curve is almost always the same. Pass rate climbs steeply from the lowest level, then flattens. The interesting question is where it flattens for your task, because everything you spend past that point buys nothing. Our walkthrough on estimating tokens for an AI task covers the arithmetic side of the same problem.

Sensible defaults by task type

From the shape of that curve, some rules of thumb that hold up across providers:

• Classification, extraction, routing, tagging. Lowest available effort. These tasks have a single correct answer that is present in the input. Deliberation adds cost, not accuracy.

• Drafting, rewriting, summarising. Low to medium. Quality here is driven by the prompt and the source material more than by thinking depth.

• Multi-step agent work, debugging, planning. Medium to high. This is the band where thinking genuinely changes outcomes, because the model has to hold several constraints at once.

• Proof-like problems, gnarly refactors, anything where a wrong answer is expensive to detect. Maximum, on the subset of requests that need it, not on the whole traffic stream.

The routing pattern matters more than any single choice. Send everything at low effort, detect failures, retry the failures at high. That costs less than running everything high and produces better results than running everything low.

The caching trap

One thing that catches teams out: on Anthropic's API, changing the effort configuration invalidates prompt cache breakpoints. The docs demonstrate it with a three-request example where changing a thinking budget from 4,000 to 8,000 tokens on the third request forces a full cache re-creation, cache_creation_input_tokens back to 1370 and cache_read_input_tokens to zero.

The lesson generalises. If you are running a cached conversation, pick an effort level and hold it. Dynamically tuning effort per turn inside one cached session can cost you more in cache misses than the lower effort saves. Route at the start of a conversation, not in the middle of one. Prompt caching explains why those misses are expensive.

Frequently asked questions

Does higher reasoning effort mean fewer hallucinations? Not reliably. More thinking helps on problems that decompose into steps, like arithmetic or multi-constraint planning. It does not conjure facts the model does not have. If the failure mode is a fabricated citation, effort is the wrong lever and grounding is the right one.

Is reasoning effort the same as temperature? No. Temperature controls randomness in token selection. Effort controls how much internal work happens before selection begins. They are independent, and turning one up does not compensate for the other. See what temperature is in AI for that dial.

Can I see the model's reasoning? Partially, and it varies. Providers typically expose summarised thinking rather than raw internal traces, and some encrypt it. Assume you can see the token count reliably and the content only sometimes.

Which level should I default to if I have no data? Whatever the provider defaults to, then run the measurement above within your first week. The defaults are reasonable middles chosen by people who cannot see your workload. They are a starting point, not an answer.

Does effort affect the maximum output length? Indirectly. Thinking tokens count against the same output ceiling as the visible answer on some APIs, so a high effort setting with a tight output limit can truncate the response. Give yourself headroom on both.

Sources checked 15 August 2026: OpenAI's reasoning guide, Anthropic's extended thinking documentation, and DeepSeek's V4-Pro GA release note. If you are trying to work out which model to point any of this at, which AI model to use for which task is the companion piece.