What is a reasoning model?

A reasoning model works through a problem in intermediate reasoning tokens before it writes its final answer. This improves results on multi-step tasks such as math, code and planning, uses more output tokens and time, and settings such as reasoning effort control it.

Concepts

How a reasoning model works

A standard language model starts writing its answer immediately. A reasoning model first generates a sequence of intermediate steps, often called thinking, and then writes the answer. During that phase it can break a problem into parts, work through calculations, test an approach and correct itself before it commits to a result.

Reasoning is a capability rather than a separate type of model. A text model can also be a reasoning model, and many let you turn reasoning on or off for each request. Some models always reason, and on those you can usually adjust how much.

When reasoning helps

Reasoning improves results on tasks with several dependent steps:

  • math, quantitative analysis and logic problems;
  • writing, reviewing and debugging code;
  • planning, and agents that choose between tools over several turns;
  • analysis that weighs long or conflicting information.

It adds little to simple lookups, classification, short rewrites or casual conversation, where it mostly adds latency and output tokens. For those tasks, lower the effort or turn reasoning off.

Prompts for reasoning models work best when they state the goal, the constraints and the expected format of the answer, and leave the planning to the model. Detailed step-by-step instructions are usually unnecessary, because working out the steps is what the reasoning phase does.

Controlling how much a model reasons

EmpirioLabs AI exposes three controls, and each model accepts the ones listed for it on this page:

  • reasoning_effort sets the depth with named levels such as low, medium and high. On many models, none turns reasoning off. The available levels differ by model.
  • thinking_budget sets the most tokens the model may spend on reasoning when thinking is on. When you leave it out, many models decide for themselves how long to reason.
  • enable_thinking switches reasoning on or off on models that offer the choice.

These parameters are accepted on /v1/chat/completions, /v1/responses and /v1/messages for models that support each endpoint. On /v1/messages you can also send a thinking object: enabled with budget_tokens sets a budget, adaptive turns thinking on without one, and disabled turns it off. On some models, a request that uses response_format or forces a tool call answers without reasoning unless it sets enable_thinking. See model parameters across endpoints for details.

Reasoning tokens, limits and cost

Reasoning is generated text, so it is billed even when you never see it. On most models reasoning tokens bill as output tokens, and each model page states how. Some models return the reasoning text, for example in reasoning_content on Chat Completions, while others return only the answer. On /v1/responses, usage.output_tokens_details.reasoning_tokens reports how many were used. On some models, preserve_thinking sends the reasoning of earlier assistant turns back to the model on /v1/chat/completions, and that earlier reasoning bills as input tokens.

Reasoning also counts toward output limits. max_completion_tokens caps reasoning and answer together, and on some models max_tokens does too. When the limit is too low, the model can spend it on reasoning and return a short or empty answer, so leave room for both.

Higher effort also means a longer wait before the answer begins. Stream the response to show progress, and for a response that runs longer than 15 minutes, use the WebSocket mode of /v1/responses. Compare reasoning-capable text models, or try different effort levels in the Playground.

Models that accept reasoning_effort, thinking_budget, enable_thinking

65 models. Each model's page has its rate card, parameters and code.

ModelParameterRange / values
DeepSeek V4 Flash 0731 DeepSeekreasoning_effort thinking_budget enable_thinkingreasoning_effort: none, low, medium, high, max; default medium
thinking_budget: 1 to 393216; default 32768
enable_thinking: default true
DeepSeek V4 Pro DeepSeekreasoning_effort thinking_budget enable_thinkingreasoning_effort: none, low, medium, high, max; default medium
thinking_budget: 1 to 393216; default 32768
enable_thinking: default true
DeepSeek V4 Pro 0813 DeepSeekreasoning_effort thinking_budget enable_thinkingreasoning_effort: none, low, medium, high, max; default medium
thinking_budget: 1 to 393216; default 32768
enable_thinking: default true
DeepSeek V4.1 Flash 1017 DeepSeekreasoning_effort thinking_budget enable_thinkingreasoning_effort: none, low, medium, high, max; default medium
thinking_budget: 1 to 393216; default 32768
enable_thinking: default true
GLM 5.1 Z.aireasoning_effort thinking_budget enable_thinkingreasoning_effort: none, low, medium, high, max; default medium
thinking_budget: 1 to 38912; default 32768
enable_thinking: default true
Gemma 4 26B-A4B Googlereasoning_effort thinking_budget enable_thinkingreasoning_effort: none, low, medium, high, max; default medium
thinking_budget: 128 to 32768; default 4096
enable_thinking: default true
Kimi K2.6 Moonshot AIreasoning_effort thinking_budget enable_thinkingreasoning_effort: none, low, medium, high, max; default medium
thinking_budget: 1 to 81920; default 32768
enable_thinking: default true
Qwen3.5 122B-A10B Alibaba Cloudreasoning_effort thinking_budget enable_thinkingreasoning_effort: none, low, medium, high, max
thinking_budget: 1 to 81920
enable_thinking: default true
Show all 65 models
ModelParameterRange / values
Qwen3.5 27B Alibaba Cloudreasoning_effort thinking_budget enable_thinkingreasoning_effort: none, low, medium, high, max
thinking_budget: 1 to 81920
enable_thinking: default true
Qwen3.5 35B-A3B Alibaba Cloudreasoning_effort thinking_budget enable_thinkingreasoning_effort: none, low, medium, high, max
thinking_budget: 1 to 81920
enable_thinking: default true
Qwen3.5 397B-A17B Alibaba Cloudreasoning_effort thinking_budget enable_thinkingreasoning_effort: none, low, medium, high, max
thinking_budget: 1 to 81920
enable_thinking: default true
Qwen3.5 4B Alibaba Cloudreasoning_effort thinking_budget enable_thinkingreasoning_effort: none, low, medium, high, max; default medium
thinking_budget: 1024 to 32768; default 4096
enable_thinking: default true
Qwen3.5 9B Alibaba Cloudreasoning_effort thinking_budget enable_thinkingreasoning_effort: none, low, medium, high, max; default medium
thinking_budget: 1024 to 32768; default 4096
enable_thinking: default true
Qwen3.6 27B Alibaba Cloudreasoning_effort thinking_budget enable_thinkingreasoning_effort: none, low, medium, high, max
thinking_budget: 1 to 131072
enable_thinking: default true
Qwen3.6 Flash Alibaba Cloudreasoning_effort thinking_budget enable_thinkingreasoning_effort: none, low, medium, high, max; default medium
thinking_budget: 1 to 64000; default 32768
enable_thinking: default true
Qwen3.6 Plus Alibaba Cloudreasoning_effort thinking_budget enable_thinkingreasoning_effort: none, low, medium, high, max
thinking_budget: 1 to 131072
enable_thinking: default true
Qwen3.7 Flash Alibaba Cloudreasoning_effort thinking_budget enable_thinkingreasoning_effort: none, low, medium, high, max
thinking_budget: 1 to 131072
enable_thinking: default true
Qwen3.7 Max Alibaba Cloudreasoning_effort thinking_budget enable_thinkingreasoning_effort: none, low, medium, high, max
thinking_budget: 1 to 262144
enable_thinking: default true
Qwen3.7 Plus Alibaba Cloudreasoning_effort thinking_budget enable_thinkingreasoning_effort: none, low, medium, high, max
thinking_budget: 1 to 262144
enable_thinking: default true
Qwen3.8 Flash Alibaba Cloudreasoning_effort thinking_budget enable_thinkingreasoning_effort: none, low, medium, high, max
thinking_budget: 1 to 262144
enable_thinking: default true
Qwen3.8 Max Alibaba Cloudreasoning_effort thinking_budget enable_thinkingreasoning_effort: none, low, medium, high, max
thinking_budget: 1 to 262144
enable_thinking: default true
Qwen3.8 Max 0902 Alibaba Cloudreasoning_effort thinking_budget enable_thinkingreasoning_effort: none, low, medium, high, max
thinking_budget: 1 to 262144
enable_thinking: default true
Qwen3.8 Omni Flash Alibaba Cloudreasoning_effort thinking_budget enable_thinkingreasoning_effort: none, low, medium, high, max
thinking_budget: 1 to 262144
enable_thinking: default true
Qwen4 Max Alibaba Cloudreasoning_effort thinking_budget enable_thinkingreasoning_effort: low, high, xhigh; default high
thinking_budget: 1 to 262144
enable_thinking: default true
GLM 5.2 Z.aireasoning_effort enable_thinkingreasoning_effort: none, minimal, low, medium, high, xhigh, max; default max
enable_thinking: default true
Kimi K3.1 Pro Moonshot AIreasoning_effort enable_thinkingreasoning_effort: low, medium, high, max; default max
enable_thinking: default true
Nova Lite 2 Amazonreasoning_effort enable_thinkingreasoning_effort: low, medium, high; default medium
enable_thinking: default true
Qwen3.5 Flash Alibaba Cloudthinking_budget enable_thinkingthinking_budget: 1 to 81920
enable_thinking: default true
Qwen3.5 Plus Alibaba Cloudthinking_budget enable_thinkingthinking_budget: 1 to 81920
enable_thinking: default true
Qwen3.6 35B A3B Alibaba Cloudreasoning_effort enable_thinkingreasoning_effort: none, low, medium, high, max; default medium
enable_thinking: default true
Qwen3.8 27B Alibaba Cloudreasoning_effort enable_thinkingreasoning_effort: none, low, medium, xhigh; default xhigh
enable_thinking: default true
Seed 2.0 Code ByteDancereasoning_effort enable_thinkingreasoning_effort: low, medium, high; default medium
enable_thinking: default true
Seed 2.0 Lite ByteDancereasoning_effort enable_thinkingreasoning_effort: low, medium, high; default medium
enable_thinking: default true
Seed 2.0 Mini ByteDancereasoning_effort enable_thinkingreasoning_effort: low, medium, high; default medium
enable_thinking: default true
Seed 2.0 Pro ByteDancereasoning_effort enable_thinkingreasoning_effort: low, medium, high; default medium
enable_thinking: default true
Seed 2.1 Turbo ByteDancereasoning_effort enable_thinkingreasoning_effort: low, medium, high; default high
enable_thinking: default true
Fugu Max Sakana AIreasoning_efforthigh, xhigh, max; default high
Fugu Ultra v1.0 Sakana AIreasoning_efforthigh, xhigh, max; default high
Fugu Ultra v1.1 Sakana AIreasoning_efforthigh, xhigh, max; default high
Fugu Ultra v2.0 Sakana AIreasoning_efforthigh, xhigh, max; default high
GLM 4.5 Flash Z.aienable_thinkingdefault true
GLM 4.6V Flash Z.aienable_thinkingdefault true
GLM 4.6V FlashX Z.aienable_thinkingdefault true
GLM 4.7 Flash Z.aienable_thinkingdefault true
GLM 4.7 FlashX Z.aienable_thinkingdefault true
GLM 5.3 Z.aireasoning_effortlow, high, max; default max
GLM 5.3 Flash Z.aireasoning_effortlow, high, max; default max
Gemini 3.8 Live Extended Thinking Googlereasoning_effortlow, medium, high; default medium
Kimi K3 Moonshot AIreasoning_effortlow, high, max; default max
MiMo V2.6 Flash Xiaomienable_thinkingdefault true
MiMo V2.6 Pro Xiaomienable_thinkingdefault true
MiMo V2.6 Pro UltraSpeed Xiaomienable_thinkingdefault true
MiniMax M3 MiniMaxenable_thinkingdefault true
MiniMax M3.1 Flash MiniMaxreasoning_effortlow, medium, high, xhigh, max; default max
Muse Glimmer 30B Meta AIreasoning_effortnone, low, medium, high, max; default high
Muse Spark 1.1 Meta AIreasoning_effortminimal, low, medium, high, xhigh; default medium
Muse Spark 1.2 Meta AIreasoning_effortminimal, low, medium, high, xhigh; default medium
Muse Spark 1.3 Meta AIreasoning_effortminimal, low, medium, high, xhigh; default medium
Perplexity Advanced Deep Research Perplexityreasoning_effortlow, medium, high; default medium
Perplexity Deep Research Perplexityreasoning_effortlow, medium, high; default medium
Perplexity Pro Search Perplexityreasoning_effortminimal, low, medium, high; default minimal
Perplexity Sonar Reasoning Pro Perplexityreasoning_effortlow, medium, high; default medium
Step 3.5 Flash 2603 StepFunreasoning_effortlow, high; default low
Step 3.7 Flash StepFunreasoning_effortlow, medium, high; default low
Step 5 Preview StepFunreasoning_effortlow, medium, high; default low

Frequently asked questions

Are reasoning tokens billed?

Yes. Reasoning is generated text, so it is billed even when the model does not return it. On most models it bills as output tokens, and each model page states how.

Can I turn reasoning off?

On many models, yes: set enable_thinking to false or reasoning_effort to none, depending on which control the model offers. Some models always reason and only let you change the effort level.

What is the difference between reasoning_effort and thinking_budget?

reasoning_effort picks a named level of depth, such as low or high. thinking_budget sets a maximum number of tokens for reasoning when thinking is on.

Why did a reasoning model return an empty answer?

When the output limit is too low, the model can spend all of it on reasoning. Raise max_completion_tokens or max_tokens, or lower the reasoning effort, so room remains for the answer.

Ready to use better endpoints?

Check out our pricing or reach out if you want your own model deployed on our stack.