How Much Do AI Tokens Cost? GPT, Claude and Gemini API Prices Explained

Every AI chatbot, writing assistant and AI feature you build on top of GPT, Claude or Gemini is billed the same way: by the token. If you do not understand token pricing, you cannot predict your bill, and surprise AI bills are one of the most common shocks for small teams and solo builders. This guide explains what a token is, what the major models charge, and how to work out your real cost before you spend anything. When you are ready to run your own numbers, use our free AI Token Cost Calculator.

What is a token, in plain English?

A token is a small chunk of text that an AI model reads or writes, roughly three quarters of an English word. Models are priced per million tokens, with input tokens (what you send) cheaper than output tokens (what the model writes back), because generating text takes more computing power than reading it.

A short question might be 20 tokens. A page of text is roughly 500 tokens. A long document, a set of instructions, or a full chat history can reach tens of thousands of tokens, and every one of them is billed on every request. That is why two apps using the same model can have wildly different bills: the cost lives in how much text moves back and forth, not in which logo is on the model.

How much do GPT models cost?

OpenAI’s GPT family is priced in tiers, from the flagship GPT-5 at roughly $1.25 per million input tokens and $10 per million output tokens, down through GPT-5 mini at about $0.25 and $2, to GPT-5 nano at about $0.05 and $0.40. The flagship costs around 25 times more than the smallest model.

In practice, that means a request with 1,500 input tokens and 500 output tokens costs a fraction of a cent on the small models and still under a cent on the flagship. The price only starts to bite at volume: a thousand of those requests a day is real money on a flagship model and pocket change on a mini. Tiered pricing exists so you can match the model to the job, and most routine jobs do not need the flagship.

How much do Claude models cost?

Anthropic’s Claude models sit at the premium end of the market. Claude Sonnet 4.5 costs roughly $3 per million input tokens and $15 per million output tokens, while the top Opus class has historically cost around $15 per million input and $75 per million output, making it the most expensive mainstream option per token.

People pay Claude prices for quality on hard tasks: long documents, careful reasoning, coding and nuanced writing. The sensible pattern is the same as with every provider, though. Test your workload on Sonnet before reaching for Opus, because for many tasks the cheaper model is already good enough, and the saving is fivefold.

How much do Gemini models cost?

Google’s Gemini models compete hard on price and context size. Gemini 2.5 Pro costs roughly $1.25 per million input tokens and $10 per million output tokens, in line with GPT-5, while Gemini 2.5 Flash drops to about $0.30 per million input and $2.50 per million output, with an even cheaper Flash Lite tier below that.

Gemini’s other advantage is its very large context window, up to a million tokens, which lets you send whole books or codebases in one request. That is genuinely useful, but remember that a million-token request is billed as a million tokens every single time you send it. Big windows are a capability, not a discount.

What about DeepSeek and other low-cost models?

DeepSeek made headlines by pricing its models at a fraction of the US providers, roughly $0.27 per million input tokens and $1.10 per million output for its V3 class model, with even lower rates for cached input. Open-weight models you run yourself, such as Llama, remove the per-token fee entirely but replace it with server costs.

Low headline prices come with trade-offs to weigh: data handling and where your text is processed, reliability and rate limits, and how well the model handles your specific task. For non-sensitive, high-volume work the savings can be dramatic. For customer data, the total cost of a mistake can dwarf the token saving, so choose with your eyes open.

How do you work out your own AI cost?

Multiply your tokens by the model’s prices. Cost per request equals input tokens divided by a million, times the input price, plus output tokens divided by a million, times the output price. Then multiply by requests per day, by 30 for a month, or by 365 for a year.

Worked example: a tool that sends 2,000 input tokens and receives 800 output tokens, used 500 times a day. On GPT-5 mini that is about $0.0021 per request, roughly $1.05 a day and $32 a month. The same workload on Claude Opus class pricing would be over $400 a month. If you do not know your token counts yet, paste a sample prompt into our AI Token Counter, and then run the full comparison in the AI Token Cost Calculator.

Why do AI bills surprise people?

Because the meter runs on text you cannot see. System prompts, tool instructions, documents and, biggest of all, conversation history are resent with every request, so a chat that starts cheap gets more expensive with every turn. Output length is the second trap: an uncapped model answering verbosely can double a bill on its own.

The fixes are boring and effective. Shorten your standing instructions. Summarise long documents once instead of pasting them repeatedly. Summarise chat history instead of resending all of it. Cap maximum output tokens. Then route each request to the smallest model that does the job well, and save the flagship for the requests that genuinely need it.

Frequently asked questions

Is it cheaper to use ChatGPT or the API?

For one person chatting occasionally, a flat subscription is usually cheaper and simpler. The API wins when you are building a product, automating volume, or embedding AI in your own tool, because you pay only for the tokens you actually use.

Do cached tokens cost less?

Yes. Most providers discount input tokens that repeat across requests, such as a long standing prompt, sometimes by half or more. If your app sends the same instructions every time, prompt caching is one of the easiest savings available.

Which model is the best value overall?

For most everyday workloads the mini and Flash class models are the value sweet spot: good quality at a tenth or less of flagship prices. Reserve flagship models for complex reasoning, long documents and final quality passes.

How accurate are the prices in this guide?

They are close approximations of public list prices at the time of writing, provided so you can compare models and estimate budgets. Providers change prices often, so confirm the current figures on the provider’s own pricing page before committing.

Prices in this article are approximate public list prices for comparison only and change frequently; always check the provider’s own pricing page before budgeting.