Token Counter for GPT, Claude, Llama

How to use the token counter for gpt, claude, llama

Paste your prompt or input text. Token counts and approximate input cost update live for each major LLM. Exact counts use the real OpenAI tokenizer (tiktoken). Other vendors use a chars-per-token heuristic because their tokenizers aren't publicly available in JavaScript. Click "Refresh prices" to pull current rates from OpenRouter.

Formula & explanation

Token count = number of BPE tokens the model's tokenizer produces. Cost = (tokens / 1,000,000) × input price per million. Output cost is 3-5x input cost and depends on response length, so it's not estimated here.

Examples

The string "Hello, world!" is 4 tokens in GPT-4o. One "average" English word ≈ 1.3 tokens. One code token ≈ 0.5 words.

Frequently asked questions

Are these token counts exact?

For OpenAI models yes — we use the actual tiktoken encoder (o200k_base) running in your browser. For Claude, Gemini, and Llama, no public JavaScript tokenizer exists, so we use a chars-per-token heuristic calibrated against published averages (~3.5 for Claude, ~4 for Gemini, ~3.8 for Llama). Expect ±10% error on those.

Are the prices current?

The prices shown by default are pinned to 2026-05-22 when this page was last updated. Click "Refresh prices" to pull live values from OpenRouter, which aggregates current rates across providers. The page falls back to the pinned values if the network call fails.

Why does the same text produce different counts on different models?

Each model family uses its own tokenizer. Recent OpenAI models use o200k_base, which is more efficient on multilingual text than the older cl100k_base. Claude, Llama, and Gemini each have their own BPE vocabularies.

Why doesn't the cost include output tokens?

Output token count is unpredictable — it depends on how long the model decides to respond. Most APIs charge 3-5x more for output than input, so total cost is roughly (input tokens) × (input price) + (output tokens) × (3-5 × input price). Multiply this tool's number by your expected response ratio.

Where do the live prices come from?

OpenRouter (openrouter.ai), which aggregates pricing across LLM providers for routing purposes. They expose a public JSON API at /api/v1/models. We fetch it only when you click the Refresh button — the page makes zero network calls otherwise.

Does this work offline?

Yes for the token counts. The tokenizer runs entirely in your browser. The live-price refresh button needs network; without it, the page shows the pinned static prices.

Why is my Claude token count an estimate?

Anthropic does not publish their tokenizer as a runnable library. The only way to get an exact Claude token count is to call their API's count_tokens endpoint, which requires an API key and a network round-trip. The heuristic we use is accurate to within ~10% for English prose.