Reasoning tokens can make a model that looks cheap on a price list cost far more than a premium one on the same job. On the 200-row spreadsheet task, Claude Sonnet 5.5 cost 17¢ per task, about twice GPT-6.1 Sol's 8.4¢, because it spent the full 16,000-token output cap on reasoning and returned no visible answer. The spreadsheet was not hard. Some models just kept thinking until they were cut off.
What I measured
I ran two tasks: analyzing a 200-row spreadsheet and reviewing 500 lines of code (CPython's textwrap.py, 497 lines). The median cost across all 326 models is 2.1¢ for the spreadsheet and 1.4¢ for the code review. Reasoning tokens are billed as output tokens, so every token spent thinking shows up on the invoice.
Twelve models were measured directly: one run per task, default settings, via OpenRouter, with output capped at 16,000 tokens. The rest are estimates. Every row below shows its basis, and capped runs are flagged. See the methodology for details.
The spreadsheet task: the cap is the cost
GPT-6 Luna was the cheapest measured model at 0.19¢ per task ($1.87 per 1,000). It used 2,338 reasoning tokens and finished. Several others used their whole budget on reasoning.
| # | 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 | |
| 9 |  Gemini 3.1 Flash LiteGoogle | 1.4¢$13.75 / 1,000 | - | 6,304 / 8,118 | estimated | |
| 10 |  MiMo-V2.6-ProXiaomi | ≥1.6¢$16.09 / 1,000 | 1479 | 4,984 / 16,001 | estimated best value capped |
The most telling comparison is between two models with the same list price of $2.00 input and $10.00 output per 1M tokens:
| Model | Reasoning tokens | Output tokens | Cost per task | Per 1,000 tasks |
|---|---|---|---|---|
| GPT-6.1 Sol | 6,553 | 919 | 8.4¢ | $84.28 |
| Claude Sonnet 5.5 | 16,000 (hit cap) | 0 | 17¢ | $170.87 |
Same price per token, about double the bill. Sonnet also has a heavier tokenizer: Anthropic models use 1.39 times as many tokens as OpenAI for the same English document, which shows in its 5,433 input tokens versus 4,779 for Sol.
Claude Haiku 5.5 shows the cheap version of the same pattern. It hit the cap with 15,434 reasoning tokens and 566 output tokens, at 0.85¢ ($8.54 per 1,000). That is still low per call, yet it is more than four times GPT-6 Luna's cost, and nearly all of it went to thinking.
The premium tier shows how large the penalty gets when a pricey model hits the cap:
- Claude Opus 5.5: 34¢ per task, 16,000 reasoning tokens, 0 output (estimated)
- Claude Fable 5.1: 85¢ per task, 16,000 reasoning tokens, 0 output (estimated)
- Qwen3.8 Max Prime: 22¢ per task, 16,000 reasoning tokens, 0 output (estimated)
Reasoning can also run past 16,000 without a cap flag. Grok 4.7 logged 30,023 reasoning tokens on this task and cost 20¢ ($199.74 per 1,000). Grok 4.3 logged the same reasoning count at 8.6¢ (estimated).
Code review shows the same pattern
GPT-6 Luna was again the lowest measured model at 0.31¢ ($3.08 per 1,000), with 4,921 reasoning tokens and 337 output tokens.
Runs that hit the 16,000-token cap:
- Qwen3.8 Flash: 0.83¢, 16,000 reasoning tokens, 0 output (measured)
- GLM 5.3 Flash: 0.87¢, 15,998 reasoning tokens, 2 output tokens (measured)
- Llama 4 Maverick: 1.1¢, 16,000 reasoning tokens, 0 output (estimated)
- DeepSeek V4.1 Flash: 2.1¢, 16,000 reasoning tokens, 0 output (measured)
- DeepSeek V4 Pro 0813: 3.5¢, 16,000 reasoning tokens, 0 output (estimated)
- GLM 5.3: 7.7¢, 15,998 reasoning tokens, 2 output tokens (estimated)
- Qwen3.8 Max Prime: 21¢, 16,000 reasoning tokens, 0 output (estimated)
None of these returned a review a developer could read. They were billed for thinking and delivered nothing. Meanwhile, models with no reasoning on this task, such as MiMo-V2.6-Pro (0.44¢), MiniMax M3 (0.45¢) and Command A+ (0.53¢), produced 2,595 to 2,611 output tokens at a fraction of the cost.
Compare Claude Sonnet 5.5 at 8.0¢, which finished with 3,476 reasoning tokens and 3,067 output tokens, with Qwen3.8 Max Prime at 21¢ for no output at all.
Which models hit the cap
- Both tasks: Qwen3.8 Flash, Qwen3.8 Max Prime, GLM 5.3 Flash, GLM 5.3, DeepSeek V4.1 Flash, DeepSeek V4 Pro 0813, Llama 4 Maverick
- Spreadsheet only: Claude Haiku 5.5, Claude Sonnet 5.5, Claude Opus 5.5, Claude Fable 5.1, MiniMax M3, MiMo-V2.6-Pro, Command A+
Reasoning is not fixed per model. It changes by task. Claude Sonnet 5.5 capped out on the spreadsheet but used 3,476 reasoning tokens on the code review.
What it means for budgeting
- Look at reasoning tokens, not just price per token. Identical list prices produced a 2x gap in cost per task.
- Treat a capped run as a failed task. It costs money and often returns no visible output.
- Test on your own inputs. My figures use one run per task, and 12 of the models are measured while the rest are estimated.
- Account for tokenizers. The same document costs more tokens on some providers.
- Budget per 1,000 tasks, then scale to your volume. The calculator accepts your own token counts.
- Cap or lower reasoning where the provider allows it. Many models here reason by default or always, so check the settings.
Full per-model breakdowns are on the spreadsheet analysis and code review pages.
FAQ
Why did Claude Sonnet 5.5 cost so much on the spreadsheet task?
It used all 16,000 tokens of the output cap on reasoning and produced no visible output. At 17¢ per task, it cost about twice GPT-6.1 Sol's 8.4¢, despite the same list price.
Is a lower price per token always cheaper?
No. Cost depends on total tokens, including reasoning tokens billed as output. Claude Haiku 5.5 has a low list price but still cost 0.85¢ on the spreadsheet because it spent 15,434 tokens reasoning.
Should I cap reasoning tokens to save money?
A cap limits the damage, but a model cut off mid-reasoning may return nothing useful. Test your own task with and without a limit before setting one in production.