# What reasoning tokens cost

A reasoning model thinks before it answers. Those thinking tokens are generated like any other output and billed at the output price even when the API hides them, so a short visible answer can carry thousands of invisible tokens. The visible answer length tells you little about the bill: the usage field of the response (output tokens) is the number to budget with.

Of the models we track, 108 support reasoning. Most charge the same output price whether or not the model reasons; [Qwen Turbo](https://apicostcheck.com/models/qwen-turbo/) list a separate, higher thinking-mode output price, which the [calculator](https://apicostcheck.com/calculator/?mode=thinking) applies when the output price switch is set to thinking. For every other model, the lever is simply how many output tokens the model produces per answer. 

## Cost per 1,000 answers of 500 visible tokens, by hidden reasoning tokens

| Model | Provider | Output /1M | +0 hidden | +2,000 hidden | +5,000 hidden |
| --- | --- | --- | --- | --- | --- |
| [MiMo V2.6 Flash](https://apicostcheck.com/models/mimo-v2.6-flash/) | Xiaomi | $0.28 | $0.14 | $0.70 | $1.54 |
| [Qwen3.8 Flash](https://apicostcheck.com/models/qwen3.8-flash/) | Alibaba (Qwen) | $0.47 | $0.23 | $1.18 | $2.58 |
| [Claude Haiku 5.5](https://apicostcheck.com/models/claude-haiku-5-5/) | Anthropic | $0.50 | $0.25 | $1.25 | $2.75 |
| [GPT-6 Luna](https://apicostcheck.com/models/gpt-6-luna/) | OpenAI | $0.50 | $0.25 | $1.25 | $2.75 |
| [GLM-5.3 Flash](https://apicostcheck.com/models/glm-5.3-flash/) | Z.ai | $0.50 | $0.25 | $1.25 | $2.75 |
| [Mistral Small (latest)](https://apicostcheck.com/models/mistral-small-latest/) | Mistral | $0.60 | $0.30 | $1.50 | $3.30 |
| [Qwen3 VL 32b Thinking](https://apicostcheck.com/models/qwen3-vl-32b-thinking/) | Alibaba (Qwen) | $0.64 | $0.32 | $1.60 | $3.52 |
| [MiMo V2.6 Pro](https://apicostcheck.com/models/mimo-v2.6-pro/) | Xiaomi | $0.87 | $0.43 | $2.17 | $4.78 |
| [DeepSeek V4 Flash](https://apicostcheck.com/models/deepseek-v4-flash/) | DeepSeek | $1.20 | $0.60 | $3.00 | $6.60 |
| [GPT-5.6 Luna](https://apicostcheck.com/models/gpt-5.6-luna/) | OpenAI | $1.20 | $0.60 | $3.00 | $6.60 |
| [MiniMax M3](https://apicostcheck.com/models/minimax-m3/) | MiniMax | $1.20 | $0.60 | $3.00 | $6.60 |
| [MiniMax M2.5](https://apicostcheck.com/models/minimax-m2.5/) | MiniMax | $1.20 | $0.60 | $3.00 | $6.60 |
| [MiniMax M2.1](https://apicostcheck.com/models/minimax-m2.1/) | MiniMax | $1.20 | $0.60 | $3.00 | $6.60 |
| [MiniMax M2](https://apicostcheck.com/models/minimax-m2/) | MiniMax | $1.20 | $0.60 | $3.00 | $6.60 |
| [GPT-5.4 nano](https://apicostcheck.com/models/gpt-5.4-nano/) | OpenAI | $1.25 | $0.62 | $3.12 | $6.87 |
| [Gemini 3.1 Flash Lite](https://apicostcheck.com/models/gemini-3.1-flash-lite/) | Google | $1.50 | $0.75 | $3.75 | $8.25 |
| [Qwen3.7 Plus](https://apicostcheck.com/models/qwen3.7-plus/) | Alibaba (Qwen) | $1.60 | $0.80 | $4.00 | $8.80 |
| [Grok Build 0.1](https://apicostcheck.com/models/grok-build-0.1/) | xAI | $2.00 | $1.00 | $5.00 | $11.00 |
| [Mistral Large 4](https://apicostcheck.com/models/mistral-large-4/) | Mistral | $2.09 | $1.04 | $5.22 | $11.49 |
| [Doubao Seed 2.1 Turbo 260628](https://apicostcheck.com/models/doubao-seed-2-1-turbo-260628/) | ByteDance | $2.15625 | $1.08 | $5.39 | $11.86 |

Estimates from published per-token rates; check official pricing before budgeting. Data synced from the LiteLLM price list 2026-10-12 12:00 UTC.

## How to keep reasoning cheap

Use the lowest reasoning effort or a non-reasoning model for routine requests and reserve deep reasoning for the questions that fail a quality check. Cap output tokens, log the output token count of real traffic, and put that average into the calculator rather than the visible answer length. Count prompts with the [token counter](https://apicostcheck.com/token-counter/); for multi-step agents, model the whole task with the [agent cost calculator](https://apicostcheck.com/agent-cost-calculator/).

## FAQ

### Are reasoning tokens returned in the response?

Usually not, or only as a summary. They are still counted in the output token usage and billed. Some providers expose a setting to show or hide them; the bill is the same either way.

### Why does my bill differ from this table?

The table fixes the visible answer and the hidden reasoning length; real reasoning length varies by question and effort setting, and input tokens are not included here. Put your measured input and output averages into the calculator.

### Is a non-reasoning model always cheaper?

Per answer, usually. Per correct answer, not always: a reasoning model that gets a hard question right the first time can beat a cheap model that needs retries or a human fix. Compare cost per completed task on your own prompts.

---
Source: https://apicostcheck.com/reasoning-tokens-cost/ · prices in USD per 1M tokens, synced 2026-10-12 12:00 UTC · Free to reuse with a link to https://apicostcheck.com (CC BY 4.0). Prices derived from LiteLLM's model price list, MIT License, (c) 2023 Berri AI.