Claude's tokenizer produces about 39% more tokens than OpenAI's for the same English text. That is why Claude Sonnet 5.5 cost more than GPT-6.1 Sol on a 20-page PDF summary, even though both list at $2.00 per 1M input tokens and $10.00 per 1M output tokens. Compare cost per task, not list price per token.
Why Claude counts more tokens
A tokenizer decides how a document is split into billable units. Two models can charge the same price per token and still bill different amounts for the same document, because one needs more tokens to represent it. The ratios below show tokens for the same English document, with OpenAI set to 1.00.
| Provider | Tokenizer ratio |
|---|---|
| Anthropic | 1.39 |
| xAI | 1.07 |
| Mistral | 1.03 |
| Qwen | 1.01 |
| Moonshot AI (Kimi) | 1.01 |
| OpenAI | 1.00 |
| Z.ai | 1.00 |
| DeepSeek | 1.00 |
| Meta (Llama) | 1.00 |
| MiniMax | 0.99 |
| 0.98 |
Anthropic is the outlier. Every other provider falls between 0.98 and 1.07. Anthropic sits at 1.39, which is the 39% in the headline.
Same list price, different bill
Sonnet 5.5 and GPT-6.1 Sol have identical list prices. The difference shows up in token counts for the same job. This is the summarize-a-20-page-PDF run, from my measured data:
| Model | List price (input / output per 1M) | Input tokens | Output tokens | Reasoning tokens | Cost per task | Cost per 1,000 tasks |
|---|---|---|---|---|---|---|
| GPT-6.1 Sol | $2.00 / $10.00 | 12,987 | 482 | 26 | 3.1¢ | $31.05 |
| Claude Sonnet 5.5 | $2.00 / $10.00 | 18,004 | 857 | 0 | 4.5¢ | $44.58 |
Sonnet 5.5 read 18,004 input tokens to Sol's 12,987, which matches the 1.39 ratio. Its output count was also higher (857 against 482), so the input gap is not the whole story. It is the part you can predict before you run anything.
The median for this task across all 326 models is 0.83¢. Both Sonnet 5.5 and Sol sit above it.
Summarizing a 20-page PDF
The input was 10,023 words of J. S. Mill's On Liberty. Here are the cheapest models for this task:
| # | Model | Cost per task | Text score | Tokens in / out | Basis | |
|---|---|---|---|---|---|---|
| 1 |  GLM 5.3 FlashZ.ai | 0.10¢$0.99 / 1,000 | 1474 | 12,995 / 1,814 | measured best value | |
| 2 |  DeepSeek V4.1 FlashDeepSeek | 0.12¢$1.22 / 1,000 | 1475 | 12,970 / 1,992 | measured best value | |
| 3 |  GPT-6 LunaOpenAI | 0.16¢$1.57 / 1,000 | 1443 | 12,987 / 533 | measured | |
| 4 |  Claude Haiku 5.5Anthropic | 0.22¢$2.20 / 1,000 | - | 18,004 / 790 | measured | |
| 5 |  Qwen3.8 FlashAlibaba Qwen | 0.32¢$3.21 / 1,000 | - | 13,119 / 2,642 | measured |
Two comparisons from the same run:
- Haiku 5.5 vs GPT-6 Luna. Both list at $0.10 per 1M input and $0.50 per 1M output. Haiku costs 0.22¢ per task and Luna 0.16¢. Haiku read 18,004 input tokens to Luna's 12,987, so the tokenizer gap shows up at the cheap end too.
- Opus 5.5 vs Sonnet 5.5. Opus lists at $4.00 / $20.00, double Sonnet's rate. It costs 8.9¢ per task against Sonnet's 4.5¢. Token counts are the same in this test (18,004 in, 857 out), so the difference comes from price alone. Note that the Opus row is estimated, not measured.
See the PDF summary task page for the full ranking.
Answering a chat question
Chat prompts are short, so input is a small share of the bill and output length dominates:
- Claude Sonnet 5.5: 1.3¢ per task ($12.68 per 1,000), with 61 input, 1,256 output and 0 reasoning tokens.
- GPT-6.1 Sol: 0.67¢ per task ($6.72 per 1,000), with 45 input, 560 output and 103 reasoning tokens.
Sonnet costs roughly 1.9 times as much as Sol here. Most of that gap comes from the output tokens Sonnet produced in this run, not from the tokenizer. The median for this task is 0.24¢.
Why list price per token misleads
List price is a rate. Your bill is the rate times the tokens, and the token count varies in three ways:
- Tokenizers differ. Claude's 1.39 ratio means a document costs 39% more in tokens before any model behavior comes into play.
- Output length differs. On the chat task, Sonnet 5.5 produced 1,256 output tokens and Sol 560.
- Reasoning tokens are billed as output. DeepSeek V4.1 Flash used 5,461 reasoning tokens on the chat question and cost 0.76¢ per task, more than its $0.30 / $1.20 list price might suggest.
Kimi K3 shows the effect from another angle. Its input lists at $0.80 per 1M, but its output lists at $15.00 per 1M. Its summarize cost is 4.3¢ per task (estimated), close to Sonnet 5.5's 4.5¢.
What to do
- Run your own document through the model you plan to use and check the input token count it reports. Do not borrow another provider's count.
- Compare cost per task, then multiply by monthly volume. The calculator does this.
- Check output length, not just the output rate. A verbose model can cost more than a higher-priced model that answers briefly.
- For Claude models, budget about 39% more input tokens than an OpenAI count of the same English text.
- Read the methodology to see which results are measured and which are estimated.
How I measured
- Measured models ran once per task at default settings via OpenRouter, with output capped at 16,000 tokens.
- Summarize input: 10,023 words of J. S. Mill's On Liberty.
- Prices are list prices as of 2026-10-11, in USD.
- Rows marked estimated are not direct measurements. Sonnet 5.5 and GPT-6.1 Sol are both measured.
You can compare the two models directly on Claude Sonnet 5.5 and GPT-6.1 Sol.
FAQ
Does Claude always cost more than GPT for the same job?
No. Token count is one factor. In my summarize test Sonnet 5.5 cost more than GPT-6.1 Sol, and in the chat test the gap came mostly from output length. Measure your own workload.
Why is Claude's tokenizer ratio 1.39?
Anthropic's tokenizer produces 1.39 times as many tokens as OpenAI's for the same English document. I do not have data on why. Every other provider in my set falls between 0.98 and 1.07.
Do reasoning tokens count toward the bill?
Yes. Reasoning tokens are billed as output tokens. DeepSeek V4.1 Flash used 5,461 reasoning tokens on my chat question, which raised its cost per task.