Overrun tells you roughly what a message to an AI (a computer program that reads your text and writes text back) will cost before you send it. After you send it, Overrun checks what it really cost and explains any difference between the two. It is free, and it runs in your web browser.
An AI model is a computer program that you send text to, and that writes text back. GPT, Gemini and Llama are names of well-known AI models. Picture a translator at a desk: you hand over a note, and a reply comes back.
A prompt is the message you send to an AI model. It can be a question, an instruction, or a long document you want summarised. Some apps also attach standing instructions to every message, such as “answer in one word”; that part is called the system prompt. If the prompt is the letter you post, the system prompt is the note clipped to every letter.
A provider is the company that runs an AI model on its computers and bills you for using it. Overrun's measurements used two providers: Groq and Google. It works like a phone company: you use their network, and they send the bill.
An AI model does not read text letter by letter or word by word. It reads text cut into small pieces called tokens: sometimes a whole word, sometimes part of a word, a space or a punctuation mark. Providers count and charge by tokens, and they price the tokens you send (input tokens) separately from the tokens the model writes back (output tokens). Think of a taxi meter that charges by the segment, not by the trip.
A tokenizer is the tool that cuts text into tokens: a pair of scissors with its own cutting pattern. Different models use different tokenizers, so the same text can come out as a different number of tokens. Overrun has four: two used by the models of OpenAI (the company behind GPT), named o200k_base and cl100k_base, one for Llama 3.1, and one for Gemma 3. Two people cutting the same ribbon with different stencils end up with a different number of pieces.
A rate card is a provider's price list. Prices are given in dollars per million tokens, because one token costs a tiny fraction of a cent. For example, Groq's price for the model gpt-oss-20b, checked on 27 Oct 2026, was $0.075 for every million input tokens. Prices change, so Overrun always shows the date each price was checked and where it came from, like the date printed on a café menu.
Counting tokens properly needs a tokenizer, so many people use a shortcut: count the characters (letters, digits, spaces, punctuation marks and emoji) and divide by four. The rule is close for English news text but goes badly wrong for many other languages, and for code and data. It is like guessing a taxi fare from the straight-line distance on a map: fine on a grid of straight streets, wrong on a winding mountain road.
Providers add tokens of their own to every message before the model reads it. You never type them, but you pay for them. Providers do not list them, so Overrun measured how many there are: in its tests, gpt-oss-20b on Groq added 71 tokens to every message, and gemini-3.1-flash-lite on Google added 1. It works like the booking fee added to every concert ticket.
An estimate is a guess you make before you send a message. The actual number is what the provider reports afterwards, in a record called a usage log. Overrun always labels the guess as an estimate, so you never mistake it for the real number. It is the builder's quote before the job, and the invoice after it.
A budget is a spending limit you set. Overrun can go through your messages one at a time, keep a running total (the ledger), and block any message whose estimated cost would take the total over the limit. It is like a prepaid card that is declined once the balance runs out.
Take the same 2,478 characters of Tamil. The rule guesses 627 tokens. GPT-4's tokenizer, cl100k_base, counts 3,352: 5.35 times the guess. A newer tokenizer, o200k_base, counts 884.
If you budget with the rule, you'd underestimate this bill by more than five times.
For comparison, the rule guessed 557 tokens for the English versions of the same sentences, and o200k_base counted 470.
The receipt above is the real one for the Tamil sample and gpt-oss-20b, built from the study's numbers.
The Findings page shows how many characters fit in one token for each kind of text, on each tokenizer. That number is called characters per token.
A usage log is the record a provider gives you, after you send messages, of how many tokens each one really used. The Reconcile page reads a usage log and compares it with the four-characters estimate, one message (a “request”) at a time. It opens with a demo log of 50 real requests that Overrun sent while measuring, so it works without any upload.
Two real cards from the demo log, priced with the groq rate card for openai/gpt-oss-20b, checked on 27 Oct 2026.
No. Nothing leaves your browser until you send it to a provider yourself. Overrun does its counting on your own computer. The only things it downloads are the tokenizers themselves: two of them come from a website called Hugging Face the first time you choose them, and your text is not sent with that download.
Each tokenizer has its own fixed list of pieces. A piece that one tokenizer keeps whole, another may have to cut into several. Tamil shows how big the difference can be: the same text is 3,352 tokens on cl100k_base and 884 on o200k_base.
When a tokenizer's list holds few pieces for a language, it has to cut that language's words into many small pieces, so the same sentence becomes more tokens. In Overrun's study, Hindi on cl100k_base needed 3.87 times as many tokens as the four-characters rule guessed.
After you send a message, trust the provider's own usage log: that is what you are billed for. Before you send, the best guess is the billed input on the receipt, which uses the tokenizer that matches your model plus the measured overhead. The four-characters estimate is the roughest of the three.
No. There is nothing to sign up for and no key to enter.