Reasoning models explained: thinking modes, effort and cost

What a reasoning model does differently

A standard language model writes its answer in a single pass, with no room for scratch work. A reasoning model first generates an internal working-out. Anthropic describes it as the model restating the question, trying approaches, checking intermediate results and dropping paths that don't hold up before it writes the final answer ^[1]. Mistral calls the same idea test-time computation: the model uses extra compute time to generate more tokens and improve its answer ^[2].

The approach grew out of prompting research. A 2022 paper showed that adding worked examples with intermediate steps to a prompt, known as chain-of-thought prompting, improved large models on arithmetic, commonsense and symbolic reasoning tasks ^[3]. Today's reasoning models are trained to produce such steps on their own, which changes how you prompt them; prompt engineering basics covers what to do differently.

How much of the thinking you get to see depends on the provider:

• A summary or nothing. OpenAI doesn't expose reasoning tokens but can return a summary on request ^[4]. Gemini returns only the final answer unless you turn on thought summaries ^[5]. On current Claude models the thinking text is omitted by default, and what you can opt into is a summary, never the raw reasoning ^[1]. xAI returns summaries for Grok 4.7 and the reasoning itself only in encrypted form ^[6].
• Redacted. Amazon Nova 2 Lite returns its reasoning as [REDACTED] ^[7].
• Full text. DeepSeek returns the chain of thought in a separate reasoning_content field ^[8].

Hiding the thinking saves time, not money: Anthropic notes that you are charged for the full thinking tokens either way ^[1].

Native, optional or none: how models differ

Our catalog records reasoning with three values. None means the model has no extended reasoning mode. Optional means reasoning can be switched on or off per request. Native means it is always on: you can usually turn its depth down, but not off. The value describes control, not quality. A switchable model with reasoning on can think as hard as a native one.

Optional does not mean off by default. Many switchable models start with thinking on, so a request that sets nothing still pays for reasoning. A selection of current models, as of October 2026:

Defaults and settings come from each provider's documentation ^{[1] [9] [4] [10] [5] [6] [8] [11] [2] [7]}. Three patterns stand out:

• Same provider, different rules. GPT-6 Sol accepts an effort of none ^[10]; GPT-6.1 Sol supports neither none nor minimal ^[4]. Among Anthropic's 5.5 models, only Claude Haiku 5.5 accepts thinking type disabled. Claude Sonnet 5.5 can drop up-front thinking with a separate setting, and Opus 5.5, like Fable 5.1, always thinks ^[1].
• Separate models instead of a switch. xAI publishes Grok 4.20 (Reasoning) and Grok 4.20 (Non-Reasoning) as separate models ^{[12] [13]}. You choose by model ID; as of October 2026 both have the same list price.
• Different names for the same dial. OpenAI uses reasoning.effort, Anthropic effort, Google thinking_level, Alibaba enable_thinking and thinking_budget, and Amazon reasoningConfig. The level names don't map one to one, so "high" at one provider tells you nothing about "high" at another.

Always on also doesn't mean maximum thinking on every request. Claude's adaptive thinking decides per request whether and how deeply to think, and at lower effort it can skip thinking on simple problems ^[9]. Gemini likewise adjusts its thinking to the complexity of the request ^[5].

What reasoning costs

OpenAI, Anthropic, Google, Amazon and Alibaba bill thinking tokens as output tokens ^{[4] [1] [5] [7] [11]}. xAI bills them as part of total usage and reports them separately ^[6]. Billing covers the full reasoning, not the summary you may receive ^{[1] [5]}. Since output is the expensive side of most price lists, as AI token pricing explained shows, thinking tokens can quickly dominate a bill.

A worked example shows the scale. GPT-6 Sol lists $2 per million input tokens and $10 per million output tokens as of October 2026. Take a request with 3,000 input tokens and a 500-token answer:

• With effort none: $0.006 for input plus $0.005 for output, $0.011 in total.
• With 4,000 thinking tokens: output grows to 4,500 tokens, so $0.006 plus $0.045, $0.051 in total.

The visible answer has the same length, but the request costs more than four and a half times as much. Over a million requests, that is $11,000 versus $51,000. The 4,000 thinking tokens are an assumption for illustration. Real counts vary with the task and the setting, so measure them: OpenAI reports them in usage.output_tokens_details.reasoning_tokens ^[4], Google in total_thought_tokens ^[5] and xAI as reasoning_tokens ^[6]. The calculator below is prefilled with that workload, 3 billion input and 4.5 billion output tokens; replace the output total with your own estimate.

Estimate the monthly API cost

Input tokens per month

Output tokens per month

Compare costs Open the calculator

This is also why list prices alone can mislead. GPT-6 Sol and GPT-6.1 Sol list the same input and output prices as of October 2026, but only GPT-6 Sol can answer without reasoning. For short, high-volume tasks, the model with the off switch can cost several times less per request at identical rates. Pricing by mode is a separate question: Alibaba notes that some hybrid models charge different prices in thinking and non-thinking mode ^[11], while Qwen3.8-Max lists one output price for both (as of October 2026).

Current input and output prices of the models in the table above:

Two further costs are easy to miss:

• Cut-off answers you still pay for. The output limit covers reasoning and answer together. If the model reaches it while thinking, OpenAI marks the response incomplete, possibly before any visible text, and still bills the input and reasoning tokens ^[4]. Google describes the same behavior ^[5]. OpenAI suggests reserving at least 25,000 tokens for reasoning and output while you experiment ^[4]. Context windows explained covers output limits in more detail.
• Context and caching. In tool-use loops, reasoning may have to be passed back to the model: in requests that carry tools, DeepSeek requires the earlier reasoning_content and returns an error if it is missing ^[8]. On current Claude models, earlier thinking stays in the conversation and is billed as input on later turns. Anthropic also notes that changing the thinking mode or top-level effort starts the prompt cache over ^[1], so switching settings mid-conversation can raise input costs.

What reasoning costs in time

Thinking happens before the answer, so it adds directly to the wait for the first visible token, and deeper settings add more:

• xAI says its highest effort level comes with correspondingly higher latency, and its code examples set client timeouts of an hour ^[6].
• Alibaba's troubleshooting notes say that turning off thinking on Qwen3.7-Plus can cut total response time by 60 to 75 percent, and that with long prompts it can bring processing time down from about 50 to about 30 seconds ^[11].
• Google advises lowering the thinking level rather than the output limit to save cost and time, because a tight limit can cut off the answer ^[5].

For interactive use, that delay often matters more than the price. For batch jobs, where nobody waits on a single response, the token count is usually the bigger concern.

When reasoning helps, and when it doesn't

The providers' guidance is consistent. Reasoning helps where the answer depends on intermediate steps:

• Math, multistep logic, complex coding and debugging ^{[1] [5] [6]}
• Planning and agentic workflows that coordinate several tools ^{[4] [7]}
• Ambiguous instructions and relationships across dense documents such as contracts or financial statements ^[14]
• Reviewing code and judging other models' answers ^[14]

It adds cost and delay with little benefit for:

• Fact retrieval and classification, where Google recommends minimal or low thinking ^[5]
• Voice applications, fast information retrieval and classification, which OpenAI lists as common uses of effort none ^[4]
• Casual chat and simple questions, where Alibaba suggests turning thinking off ^[11]

More effort isn't automatically better. In its guidance for Claude Opus 4.7, Anthropic writes that the top effort level adds significant cost for relatively small quality gains on most workloads and can lead to overthinking on some structured-output tasks ^[9].

Many workflows don't need to choose one mode. OpenAI's reasoning best-practices guide describes reasoning models as planners and non-reasoning models as workhorses, and expects most workflows to combine them: a reasoning model plans or decides, faster models carry out the individual steps ^[14]. With a switchable model, you can do the same inside one model by setting effort per request.

How to choose a setting

1. Build a small test set from real tasks with a clear pass criterion. How to choose an AI model walks through this.
2. Start at the default or one step below, then sweep. Anthropic recommends running an effort sweep on your own evaluations rather than carrying settings over from an earlier model ^[9].
3. Compare across models, not only across levels. A smaller model at a high effort level is worth testing against a larger model at a lower one. Anthropic suggests comparing Claude Haiku 5.5 at its top levels with Claude Sonnet 5.5 on quality, cost and speed ^[9].
4. Count the thinking tokens in the usage data and enter the real output totals in the AI cost calculator.
5. Steer with effort, cap with the output limit. Anthropic calls effort soft guidance and the output limit the strict ceiling on spend ^[1]. Leave enough headroom that the cap doesn't cut off answers.
6. Keep settings stable in cached conversations. On Claude, changing the top-level effort between requests restarts the prompt cache; newer models accept a per-message effort change that keeps it ^[9].

The compare page shows each model's reasoning value next to its prices, and each model page notes the documented reasoning controls. AI pricing lists current output prices, which matter most for reasoning workloads. How we collect and verify this data is described in our methodology.