Claude vs GPT token count: a real comparison
On a real 882-word technical document, GPT's own tokenizer counted 1,951 tokens. The tokenizer used to approximate Claude and Gemini counted 1,948, a difference of 0.15%. DeepSeek's real tokenizer counted 2,040. The full breakdown, and why two of these six numbers are approximations rather than facts, is below.
The test document
To avoid making up numbers, we ran this project's own architecture guide (882 words, 7,694 characters of real technical prose, not marketing copy) through NoRefund's CLI against six models from six different providers. Here is exactly what came back.
| Model | Provider | Tokenizer used | Tokens | Exact or approximate |
|---|---|---|---|---|
| GPT-5.6 Sol | OpenAI | tiktoken, gpt-4o encoding | 1,951 | Exact |
| Claude Sonnet 5 | Anthropic | cl100k_base (stand-in) | 1,948 | Approximate |
| Gemini 3.1 Pro Preview | cl100k_base (stand-in) | 1,948 | Approximate | |
| DeepSeek V3 | DeepSeek | HuggingFace, real | 2,040 | Exact |
| Qwen2.5 72B | Qwen | HuggingFace, real | 1,948 | Exact |
| Mistral Small 24B | Mistral | HuggingFace, real | 1,999 | Exact |
Generated with norefund <file> --model <id> against each model, on 2026-08-26. Re-derivable at any time; see Data Sources for the verification method behind each tokenizer mapping.
Why Claude and Gemini tie exactly
That is not a coincidence or a bug. Anthropic and Google are both among the handful of providers that keep their production tokenizer private. NoRefund, like most tools in this space, falls back to cl100k_base as the closest public equivalent for both, and labels both "approx." in the Model Registry. When two models share the same fallback tokenizer, they will always produce the identical count for a given document, because it genuinely is the same tokenizer running twice.
The gap between that approximation and GPT's real count was 0.15% on this document. That is a fairly plain, English-language technical guide. Documents with heavy code blocks, tables, non-English text, or unusual Unicode can widen that gap further, since tokenizers diverge more on content their training data saw less of.
What this means for choosing a model
If you are optimizing cost across providers, the token count is only half the equation, since Claude, Gemini and GPT also charge different per-token prices. See the LLM cost calculator for how the same document turns into different dollar amounts once pricing is factored in, or run your own document through NoRefund's Compare Models screen to get a ranked, cheapest-first answer for your specific file.