Find out what your prompt actually costs.

About 750 words of ordinary English, or roughly 4,000 characters — around two and a half pages of double-spaced text. That works out to about 1.3 tokens per word. Technical writing, code, tables and non-English text all use more tokens for the same word count. The reference table above converts the common sizes.

It's a close estimate, not the provider's own tokeniser. The page models the same rules real byte-pair tokenisers follow — whole common words as single tokens, leading spaces attached to words, long words split into fragments, digits grouped, punctuation separate — then applies a per-family adjustment. On ordinary English prose it typically lands within a few percent. Code, unusual formatting and non-Latin scripts drift further. For billing-exact numbers, read the usage field the API returns with every response.

No. Everything runs locally in your browser. There's no request to a server, no analytics on the content, and nothing stored when you close the tab. You can disconnect from the network and the calculator still works.

Two reasons. Each model family has its own vocabulary, so the same sentence splits into a slightly different number of tokens. And each has its own price per million tokens. A model with a cheaper rate but a coarser tokeniser wins twice over.

Everything you send: the system prompt, every earlier message in the conversation, tool and function definitions, tool results, and any documents or images you attach. In a long chat the history is resent on every turn — the Conversation tab above shows how quickly that compounds.

By dimensions, using a per-provider formula. OpenAI charges 85 base tokens plus 170 per 512-pixel tile at high detail. Anthropic approximates width × height ÷ 750. Google charges 258 tokens per 768-pixel tile. The same 1024×1024 image costs about 765 tokens on OpenAI, 1,400 on Anthropic and 1,032 on Google — a near-two-fold spread for identical input. The image calculator above does the arithmetic for any size.

Providers can store the unchanging prefix of your prompt — typically a long system message or a fixed document — and re-read it at a large discount, usually around 10% of the normal input rate. There's often a small premium to write the cache initially. It pays off when the same prefix is reused within minutes, which describes most production chat and retrieval systems.

The maximum number of tokens the model can hold at once, counting your input and its reply together. Exceed it and the request fails or the oldest part of the conversation is dropped. The meter in the estimate panel shows how close you are for the model you picked.

Yes — use Upload file or drag a text file onto the input box. Plain text, Markdown, JSON, CSV and source code all work. Large files may take a moment to render the visual breakdown, so the display caps at the first 1,500 tokens with an option to expand.

Reading your prompt happens in one parallel pass. Generating a reply happens one token at a time, each step requiring a full forward pass through the model, so it uses far more compute per token. Providers price that difference in — which is why capping reply length saves more than trimming prompts.

Often, yes. Models that think before answering bill their internal reasoning as output tokens, even though you never see it. A 200-word visible answer can carry several thousand reasoning tokens behind it. If you're budgeting for a reasoning model, raise your output estimate substantially rather than trusting the visible reply length.

They are published list prices for each provider's API, reviewed in May 2026 and stated in US dollars per million tokens. Open-weight models such as Llama, Mistral Small, DeepSeek and Qwen have no single official price, so the figures shown are typical rates from hosted inference providers. Every price cell is editable — enter your own contracted rate and the whole page recalculates.