ChatGPT token counter & OpenAI cost calculator

How GPT counts tokens

OpenAI's tokeniser is the one everything else gets measured against, largely because it was the first with a public implementation. It is byte-level BPE: common English words are a single token, rarer ones split into pieces, and a leading space merges into the word that follows it — which is why "cat" and " cat" cost the same but "Cat" sometimes does not.

The counter above is an estimate, not OpenAI's own merge table. It models the behaviour that actually drives the number — a leading space merging into the word after it, digits grouping, characters outside the Latin range costing several times more — which is close enough to size a prompt and budget a workload, and deliberately not close enough to reconcile an invoice against.

The spread between OpenAI models is the widest of any provider

This is the practical thing to know about the OpenAI line-up, and it does not apply to the others. The cheapest and dearest models here differ by more than an order of magnitude on input, and considerably more on output. Two models with the same context window and the same tokeniser can produce bills that are not remotely comparable. On Claude or Gemini, picking the wrong model costs you some money; on OpenAI it can cost you twenty times the money. Sort the table by cheapest and look at where your workload actually lands before assuming you need the flagship.

What the 11 GPT models cost

Take one representative call — a 2,000-token prompt and a 600-token reply, which is roughly two pages in and one page out — and run it 100 times a day.

Output is where the bill lives. Every OpenAI model here charges several times more for tokens it writes than for tokens it reads — so a prompt that doubles in length is usually cheaper than a reply that does. If you are trying to cut spend, cap the reply before you trim the prompt.

Which GPT model to use

Start at the bottom of the price list, not the top. The nano and mini tiers handle classification, extraction, routing, tagging and short rewrites at a quality most people cannot distinguish from the flagship on those jobs, for a fraction of the cost. Move up when the task needs multi-step reasoning, long-range consistency across a big document, or code that has to be right first time. The flagship earns its price on hard problems and wastes it on easy ones, and most production traffic is easy ones.

Context windows

The GPT models here run from 128K tokens to 1.05M. That window is shared: your prompt, every earlier turn in the conversation, and the reply all have to fit inside it together. The table above marks a model in red when the text you have pasted plus the reply length you set would not fit — which is usually how people discover that a window is not as generous as the headline number suggests.

Because the range spans 128K to 1.05M, a prompt that works on one GPT model can fail on another in the same family. Check the fit before you switch model to save money.

Questions

How many tokens is a ChatGPT message?

For ordinary English, roughly 1.3 tokens per word — so a 500-word message is about 650 tokens. Code runs much higher because indentation and punctuation are each charged, and non-Latin scripts higher again. Paste the real text above rather than estimating from a rule of thumb; the ratio moves more than people expect.

Does the ChatGPT web app charge tokens?

No. A ChatGPT Plus or Free subscription is a flat monthly price with usage limits, not a token bill. Tokens are billed on the API, which is what this calculator prices. The counts still matter on the web app though, because the context window is the same limit either way.

Is this the same tokeniser ChatGPT uses?

It is a close estimate, not the real merge table. This counter models the behaviour of byte-level BPE — how spaces attach, how digits group, how much more a non-Latin character costs — rather than shipping OpenAI's vocabulary. For prose it lands within a few percent. For an exact figure on a specific model, use OpenAI's own tokeniser library.

Why does my prompt cost more than I counted?

Three usual reasons. The reply counts too, and is billed at the higher output rate. Anything in the conversation above your message is resent with every turn, so a long thread pays for its own history repeatedly. And a system prompt is charged on every call even though you only wrote it once.

Counting for a different provider

The same text costs a different number of tokens on every family, so if you are comparing providers, compare on your own text rather than on a rule of thumb.

Rates are USD per million tokens, published list price, checked 2026-08-31. Batch and cached-input discounts are not applied. Confirm against OpenAI before committing a budget.