Prompt Token Cost Calculator

A token is a sub-word unit produced by the model's tokenizer. For OpenAI's o200k encoding (GPT-4o family) one token is roughly 3-4 English characters or about 0.75 of a word. Whitespace, punctuation, code, and CJK characters all tokenize differently — that's why two prompts of the same character count can have very different token counts. Pricing is always quoted per 1,000,000 tokens.

Anthropic, Google, and DeepSeek do not publish a JavaScript-runnable tokenizer. Anthropic's official advice is to use their server-side count_tokens endpoint, which would require sending your prompt to their server — defeating the purpose of an offline tool. We instead use a character heuristic (~4 chars/token English, ~2 chars/token CJK) which is accurate to within roughly 5-15% for natural-language prompts. Code, tables, and JSON tend to tokenize denser than natural language, so for those workloads add a 10-20% safety margin on top of the displayed cost.

Prompt caching lets you reuse a long system prompt or document context across many calls. Anthropic charges 1.25x the input price the first time the cache block is written (5-minute TTL) and only 0.1x on subsequent reads — so a 50K-token system prompt reused 100 times costs roughly the equivalent of 26K tokens billed at full price, instead of 5,000K tokens. OpenAI's prompt caching is automatic and bills cache reads at 0.5x the input price (no write premium). Toggle the cache option, set the cache-read share slider to your expected hit rate, and the table updates instantly.

GPT-4o ($2.50 / $10 per 1M tokens) is the cheapest of the three frontier models and ships with the lowest latency, making it ideal for chat UIs and low-stakes generation. Claude Opus 4.7 ($15 / $75) is the most expensive but tends to win on long-form reasoning, careful writing, and tool-use accuracy — pair it with prompt caching to keep the bill manageable. Gemini 2.5 Pro ($1.25 / $10) is the cheapest input-side option of the three, has a 2M-token context window, and excels at multimodal (vision, video) tasks. For high-volume background jobs (extraction, classification) consider GPT-4o mini, Claude Haiku 4.5, Gemini 2.5 Flash, or DeepSeek-V3 — each under $1/1M input tokens.

Exact. The library ships the same BPE rank tables that OpenAI uses internally and the API. The o200k_base encoder is used for GPT-4o, GPT-4.1, and the entire o-series (o1, o3, o4-mini); cl100k_base is used for GPT-4 Turbo and GPT-3.5. The number this calculator reports for an OpenAI row will match the usage.prompt_tokens field returned by the OpenAI API to the digit, modulo special tokens added by chat formatting (typically 3-7 extra tokens per message). For raw text completions the count is exact.

The price table is bundled in src/data/llm-model-pricing.ts and is purely data — no UI is needed to change it. If you fork this repo or run it locally, edit that file to add a row with your provider's pricing, set tokenizer to 'approx' (or 'cl100k_base' / 'o200k_base' if your model uses an OpenAI tokenizer), and rebuild. The lastUpdated field is there to make pricing audits easy.

Yes — completely. Tokenization runs in your browser via the gpt-tokenizer JavaScript library; no prompt text, no token counts, and no cost numbers are ever transmitted to a server. We do not log prompts, do not run analytics on input text, and do not use a remote tokenizer endpoint (which would defeat the purpose). The page works offline after the first load, so you can paste internal system prompts, customer data, or unreleased product copy without leaving any trace.