Reasoning tokens are billed as output, and they can outnumber the visible answer by a wide margin. That is why a model with a low list price can still cost more per task than its price per million tokens suggests. In my runs, DeepSeek V4.1 Flash produced 5,461 reasoning tokens for a chat answer that was 859 tokens long.
What reasoning tokens are
Reasoning models generate hidden working before the answer you see. You pay for that working at the output rate. A short answer can still be an expensive one.
Two numbers decide the bill:
- Output price per million tokens. The rate the model charges for output.
- Reasoning token count. How many hidden tokens the model generated for that task.
The price gives the rate. The count tells you how many tokens you pay it on. My task pages show both, and mark each figure as measured or estimated. See /methodology/ for details.
Case 1: a chat answer on DeepSeek V4.1 Flash
In my chatbot answer task, DeepSeek V4.1 Flash used 65 input tokens, returned 859 visible output tokens and logged 5,461 reasoning tokens. The reasoning is more than six times the visible answer.
- Billed output: 859 + 5,461 = 6,320 tokens
- Output price: $1.20 per million tokens
- Cost per task: 0.76¢, or $7.60 per 1,000 tasks
Input is negligible. Nearly all the cost is hidden thinking. Check the model page for list prices, or run your own volumes in the calculator.
Other models on the same task:
| Model (chat answer) | Visible output | Reasoning tokens | Cost per task |
|---|---|---|---|
| GPT-6 Luna | 540 | 639 | 0.06¢ |
| GPT-6.1 Sol | 560 | 103 | 0.67¢ |
| DeepSeek V4.1 Flash | 859 | 5,461 | 0.76¢ |
| Claude Sonnet 5.5 | 1,256 | 0 | 1.3¢ |
Claude Sonnet 5.5 logged no reasoning tokens, yet costs more per task than DeepSeek V4.1 Flash. Its visible output is longer and its price is higher: $2.00 input and $10.00 output per million tokens, against $0.30 and $1.20. Reasoning is one cost driver. Price per token is another.
Case 2: spreadsheet analysis and the 16,000-token cap
In the spreadsheet analysis task, the model gets a 200-row spreadsheet. Output was capped at 16,000 tokens, and many runs hit the cap. When that happens, the output is almost entirely reasoning, and the bill follows.
- Claude Sonnet 5.5 (measured): 0 visible output, 16,000 reasoning tokens, 17¢ per task, $170.87 per 1,000.
- Claude Opus 5.5 (estimated): same token profile, 34¢ per task.
- Claude Fable 5.1 (estimated): same token profile, 85¢ per task.
- Claude Haiku 5.5 (measured): 566 visible output, 15,434 reasoning, 0.85¢ per task.
The median across all 326 models on this task is 2.1¢. Claude Sonnet 5.5 at 17¢ is roughly eight times that. The cap hit all four Claude models, so price per token sets the gap between them: the same 16,000 reasoning tokens cost 0.85¢ on Haiku and 85¢ on Fable.
Grok 4.7 shows a different pattern: 1,263 visible output tokens and 30,023 reasoning tokens, for 20¢ per task ($199.74 per 1,000). Its output price of $6.00 per million is lower than Claude Sonnet 5.5's $10.00, but its reasoning volume is far higher.
| # | Model | Cost per task | Text score | Tokens in / out | Basis | |
|---|---|---|---|---|---|---|
| 1 |  GPT-6 LunaOpenAI | 0.19¢$1.87 / 1,000 | 1443 | 4,779 / 2,794 | measured best value | |
| 2 |  DeepSeek V4.1 FlashDeepSeek | ≥0.26¢$2.65 / 1,000 | 1475 | 4,801 / 16,000 | measured best value capped | |
| 3 |  GLM 5.3 FlashZ.ai | ≥0.34¢$3.44 / 1,000 | 1474 | 4,900 / 16,000 | measured capped | |
| 4 | GPT-5.4 NanoOpenAI | 0.44¢$4.45 / 1,000 | 1401 | 4,779 / 2,794 | estimated | |
| 5 |  Mistral Small 4Mistral | 0.71¢$7.14 / 1,000 | - | 4,774 / 10,700 | estimated | |
| 6 |  Qwen3.8 FlashAlibaba Qwen | ≥0.85¢$8.47 / 1,000 | - | 6,357 / 16,000 | measured capped | |
| 7 |  Claude Haiku 5.5Anthropic | ≥0.85¢$8.54 / 1,000 | - | 5,433 / 16,000 | measured capped | |
| 8 |  Llama 4 MaverickMeta | ≥1.2¢$11.62 / 1,000 | 1327 | 6,306 / 16,000 | estimated capped |
"Reasoning none" is not a guarantee
Some models are labeled "reasoning none" in my catalog. On the spreadsheet task, several still logged large reasoning counts:
- MiMo-V2.6-Pro (estimated): 15,847 reasoning tokens, 1.6¢ per task.
- MiniMax M3 (measured): 15,847 reasoning tokens, 2.1¢ per task.
- Command A+ (estimated): 15,847 reasoning tokens, 2.5¢ per task.
- Llama 4 Maverick (estimated): 16,000 reasoning tokens, 1.2¢ per task.
- Qwen3.8 Flash (measured): 16,000 reasoning tokens, 0.85¢ per task.
The label is a catalog setting, not a promise about every task. Even on the chat task, MiMo-V2.6-Pro, MiniMax M3 and Command A+ each logged 482 reasoning tokens. Do not assume a model is cheap because its label says "none."
What to do about it
- Lower the reasoning effort. Where the API exposes a reasoning setting, reduce it for simple work such as classification, extraction and short answers. This should cut reasoning tokens, but test it on your own inputs because quality can drop.
- Set an output cap. A cap limits the worst case. My spreadsheet runs show what happens at 16,000 tokens.
- Match the model to the task. GPT-6 Luna cost 0.19¢ on the spreadsheet task with 2,338 reasoning tokens, while several models burned 16,000.
- Measure, do not guess. Count reasoning tokens in your own logs and compare cost per task, not list price.
Cost per task is the output price times the tokens the model thinks and writes, plus input. Check both before choosing a model for production.
FAQ
Are reasoning tokens billed?
Yes, as output tokens. On the chat task, DeepSeek V4.1 Flash's 5,461 reasoning tokens were billed alongside its 859 visible output tokens.
Why does a model labeled "reasoning none" still produce reasoning tokens?
The label is a catalog setting, and the count varies by task. On the spreadsheet task, MiMo-V2.6-Pro and MiniMax M3 both logged 15,847 reasoning tokens.
Does a lower price per million tokens mean a cheaper task?
Not always. Claude Sonnet 5.5 logged no reasoning on the chat task and still cost 1.3¢ against DeepSeek V4.1 Flash's 0.76¢, because of its higher price and longer answer. Compare cost per task.