Paste any text to count its tokens, then see exactly what it costs to run on Claude, GPT, Gemini, Grok, DeepSeek, and 53+ other models. Everything runs in your browser — nothing you paste leaves the page.
1 token ≈ 4 characters ≈ ¾ of an English word. Cached input is the share of your prompt that repeats between requests — system prompts, few-shot examples — billed at the provider’s cache-read rate where one exists.
| # | MODEL | INPUT COST | OUTPUT COST | TOTAL | VS CHEAPEST |
|---|---|---|---|---|---|
| 1 | Granite 4.1 8B IBM · $0.05 in / $0.10 out | $0.00 | <$0.0001 | <$0.0001 | CHEAPEST |
| 2 | Granite 4.0 Micro IBM · $0.017 in / $0.112 out | $0.00 | <$0.0001 | <$0.0001 | +12% |
| 3 | Nova Micro Amazon · $0.035 in / $0.14 out | $0.00 | <$0.0001 | <$0.0001 | +40% |
| 4 | Phi 4 Microsoft · $0.07 in / $0.14 out | $0.00 | <$0.0001 | <$0.0001 | +40% |
| 5 | Command R7B Cohere · $0.037 in / $0.15 out | $0.00 | <$0.0001 | <$0.0001 | +50% |
| 6 | Nemotron 3.5 Lightning NVIDIA · $0.08 in / $0.20 out | $0.00 | $0.0001 | $0.0001 | +100% |
| 7 | MiMo-V2.5 Xiaomi · $0.12 in / $0.24 out | $0.00 | $0.00012 | $0.00012 | +140% |
| 8 | GLM-5.3-Flash Z.ai · $0.075 in / $0.25 out | $0.00 | $0.00013 | $0.00013 | +150% |
| 9 | Llama 4 Scout Meta · $0.10 in / $0.30 out | $0.00 | $0.00015 | $0.00015 | +200% |
| 10 | Nemotron 3 Super NVIDIA · $0.085 in / $0.40 out | $0.00 | $0.0002 | $0.0002 | +300% |
| 11 | Qwen3.8-Flash Alibaba · $0.15 in / $0.47 out | $0.00 | $0.00023 | $0.00023 | +370% |
| 12 | Hunyuan A13B Tencent · $0.14 in / $0.57 out | $0.00 | $0.00028 | $0.00028 | +470% |
Prices verified August 30, 2026 against official provider pricing pages. Want the raw rates? Browse the full pricing table or compare models side by side.
Before a model sees your text, a tokenizer splits it into chunks — whole words, word fragments, punctuation. You’re billed per chunk, in both directions: the prompt you send and the reply you get back, at separate rates.
Every vendor ships its own tokenizer, so the same paragraph is a slightly different count on Claude than on GPT or Gemini. The ≈4-characters-per-token heuristic this calculator uses lands within about ±10–15% for English prose — close enough to budget with. The exact count your bill reflects is in the usage field of every API response.
It uses the standard estimation heuristic — 1 token ≈ 4 characters ≈ ¾ of an English word — which lands within about ±10–15% for English prose. Code, CJK text, and emoji tokenize differently per model, since every vendor ships its own tokenizer. For billing-grade numbers, the usage field in the API response is the ground truth.
Roughly 750 English words, or about 4,000 characters — a page and a half of prose. Scaling up: 100K tokens is a short novel, and 1 million tokens is around 750,000 words.
Generating text is sequential — the model produces one token at a time, each requiring a full forward pass — while input tokens are processed in parallel. Providers price that difference at 3–6× per token, which is why verbose responses, retries, and chain-of-thought dominate most bills.
If you resend the same prompt prefix — a system prompt, few-shot examples, long documents — providers serve it from cache at roughly a tenth of the fresh-input rate. Set the percentage of your input that repeats between requests and the calculator applies each model’s cache-read rate to that share. Models without cached pricing charge the full rate regardless.
No. The counting and cost math run entirely in your browser — nothing you type or paste leaves the page, and we don’t log or store it.
Give every API key a credit balance — validate, deduct, enforce limits, and watch per-consumer AI spend in real time. One call from your own middleware.