Why is this more accurate than a local estimator?

Local estimators approximate tokenization with heuristics. This tool asks the provider itself how many tokens your text uses, so the count exactly matches what you would be billed for that model.

Where does my API key go?

Your key is held only in your browser memory for the request and sent directly to OpenAI or Anthropic over HTTPS. It is never sent to this site's servers, stored, or logged.

Which endpoints does it call?

For Anthropic it calls the messages count_tokens endpoint. For OpenAI it sends a minimal chat completion with max_tokens set low and reads prompt_tokens from the usage field, which reflects the exact prompt tokenization.

Will this cost me anything?

The Anthropic count-tokens endpoint is free. The OpenAI method issues a tiny completion to read usage, costing a negligible amount. Both use your own key and quota.

What if I get an error?

Common causes are an invalid key, a model name your key cannot access, or a network or CORS issue. The tool shows the provider's error message so you can correct the key or model.

Token Count API Tester

Token count API tester

Stop guessing token counts. This tool sends your text to OpenAI or Anthropic with your own API key and returns the exact token total the provider reports — the same number you are billed for — rather than a local heuristic estimate. It is the reliable way to verify a production prompt fits your context window and budget.

How it works

Each provider exposes the true tokenizer behind its API:

Anthropic has a dedicated count_tokens endpoint that returns input token count for a given model and message with no generation cost.
OpenAI does not expose a standalone counter, so the tool sends a minimal request with a tiny max_tokens and reads usage.prompt_tokens from the response, which is the exact tokenization of your input.

Your key never touches this site’s servers — the request goes straight from your browser to the provider over HTTPS.

Tips and notes

Match the model. Tokenization differs slightly across model families, so pick the model you will actually call.
Use it for context-window checks. Compare the returned count against the model’s context limit before shipping a long-document prompt.
Keep keys safe. Use a restricted or low-quota key for testing, and rotate it if you ever paste it into a shared machine.