← All Tools
Token Counter & Cost Estimator
Estimate the token count of a prompt and the cost of running it through the major LLM APIs. Useful for budgeting RAG pipelines, weighing model choices, and avoiding nasty bill surprises.
100% client-side. Your prompt is processed locally. Nothing is sent to any vendor or to us — estimation only.
⚠ Approximate, not exact. The page uses heuristic tokenizers calibrated against the real ones. Expect ±5% on English prose, more on code or non-Latin text. For exact billing, run the vendor's own tokenizer.
—
Output assumption
Most API costs split input from output, with output ~3-5× more expensive. Set the assumed response length:
× input tokens
tokens
Volume
Token estimates by tokenizer family
| Tokenizer family | Input | Output (est.) | Total |
|---|
Cost matrix
Pricing snapshot dated . Click a row to override the pricing for that model.
| Model | Input $/1M | Output $/1M | Per call | Volume total |
|---|
How the estimates work
- GPT-family (cl100k / o200k base): chars/4.0 baseline, plus a word-frequency bias term. ChatGPT and most OpenAI-compatible models.
- Claude: Claude's BPE is similar to cl100k for English; we use chars/3.8 as a conservative estimate (Claude tends to use slightly more tokens).
- Gemini: Gemini's SentencePiece tokenizer is competitive; we use chars/3.7.
- Llama 3 / DeepSeek / Qwen / Mistral: 128k-vocab BPE, similar density to GPT-4o. We use chars/3.9.
- CJK / non-Latin: each character often = 1-2 tokens. We add a CJK-character-count contribution.
- Code & high-symbol text: dense punctuation reduces token efficiency by 10-20%; the heuristic adds an adjustment when symbol density > 25%.