ChatGPT API cost calculator

ChatGPT API cost: how to work it out

ChatGPT API cost is billed per million tokens, pieces of text of about four characters. Input (what you send the model) and output (what it writes back) are priced separately, and output usually costs five times as much. A chatbot handling 1,000 conversations a month in English costs between EUR 1.09 and EUR 108.76 at the prices of 22 October 2026, depending on the model. The calculator above works it out for your own volumes across thirteen models from OpenAI, Anthropic, Google and Mistral, in euros, pounds or dollars at the ECB rate.

The formula, step by step, for a chatbot with four questions per conversation, 1,500 words of fixed instructions and knowledge, questions of 25 words and answers of 120:

1. Words to tokens. English takes about 1.3 tokens per word: 1,500 words become 1,950 tokens, a question 33, an answer 156.
2. Input per conversation. A model remembers nothing between two questions. Every question resends the instructions and the whole conversation: four times 1,950, plus the questions ten times (1 + 2 + 3 + 4), plus the answers six times (0 + 1 + 2 + 3). In total 9,066 tokens.
3. Output per conversation. Four answers of 156 tokens: 624 tokens.
4. Per month. Times 1,000 conversations: 9.1 million tokens in and 0.6 million out.
5. Price. On Claude Sonnet 5.5 (USD 2 in, USD 10 out per million): 18.13 plus 6.24 is USD 24.37, or EUR 21.75 at USD 1.1205 to the euro.
6. With caching. What the previous question already sent, the model reads from its cache. At Anthropic that cache lives five minutes. Within a conversation that is plenty: questions 2, 3 and 4 read the instructions and the conversation so far from the cache. Between two conversations it is not: 1,000 a month is about one every eighteen minutes, so the first question sends everything fresh. That puts 6.5 of the 9.1 million input tokens in the cache, at USD 0.10 per million. The other 2.6 million are written to the cache at 1.25 times the input price, USD 2.50. Together 6.38 plus 0.65 plus 6.24 is USD 13.27, or EUR 11.84. On GPT-6.1 Sol the cache lives at least half an hour, so the instructions stay warm between conversations too: EUR 7.66.

Prices of 22 October 2026, in US dollars per million tokens, before VAT. The last column is the chatbot above in English, without caching, in euros a month.

Two prices have announced changes: Gemini 4 Argon doubles on 1 December 2026, Gemini 3.8 Flash on 1 January 2027. The calculator shows both amounts.

Tokens in English, Dutch and other European languages

A token is a piece of text the model reads as one unit: a short word, part of a long word, a punctuation mark. OpenAI counts about four characters per token, or three quarters of an English word. Other languages cost more. Long compound words are cut into more pieces, so one measurement across six tokenizers found 1.67 tokens per Dutch word against 1.23 per English word. The same chatbot answering in Dutch costs about a third more than in English, and a site serving buyers in several languages should budget per language.

Every provider cuts text its own way. Anthropic notes that its tokenizer since Claude 4.7 produces about 30 percent more tokens for the same text, so allow more headroom on Claude. What Claude costs as a subscription rather than per token is in Claude pricing.

Four worked examples for a small or mid-sized company

The same formula applied to four things companies have built today. Euros a month, before VAT, English text. Light is GPT-6 Luna or Claude Haiku 5.5, workhorse GPT-6.1 Sol or Claude Sonnet 5.5, top GPT-6 Astra or Claude Fable 5.1.

• The chatbotof a practice, an installer or a B2B supplier: 1,000 conversations is about 33 a day, one every eighteen minutes over a ten-hour day. Too slow to keep Anthropic's cache warm between conversations, fast enough for OpenAI's; that is why the caching row differs per provider. On a light model a conversation costs a tenth of a cent.
• Drafting e-mail replies1,000 customer e-mails with 400 words of instructions and standard answers, replies of 150 words. A person checks and sends.
• Reading invoices300 invoices of 400 words, the model pulls out supplier, date, amounts and VAT. A scan sent as an image is counted in image tokens under each provider's own rules.
• Drafting quotes40 requests a month, with your price list, terms and one example quote, 2,500 words in all, as fixed knowledge. The model writes a first draft of 900 words.

E-mails and invoices can often wait until the next day. Sent as a batch, they cost half at all four providers.

Subscription or API: ChatGPT Plus and Business against paying per token

A subscription and the API are two different things. With a subscription a person uses the ChatGPT or Claude app. With the API you pay per token for software that runs on its own: a chatbot on your site, a link to your mailbox, a tool that reads invoices. Which app suits your team is in ChatGPT vs Claude and ChatGPT vs Gemini; every OpenAI plan is in ChatGPT pricing.

A team of five on ChatGPT Business pays EUR 105 a month. For the same amount a chatbot on GPT-6.1 Sol answers more than 4,800 conversations in English. They do not replace each other: the subscription makes people faster, the API lets software take work over.

VAT on your API invoice

Providers list their prices before VAT. Anthropic says your billing address decides which tax applies and that you add a VAT ID in its console. Enter your company's VAT number with every provider before the first invoice. If Belgian VAT is charged, it is 21 percent; in the Netherlands, Germany or the UK your own country's rules apply. How the invoice goes into your books is a question for your accountant.

Where the real cost sits

The tokens are almost never the big line. What costs money: building the application, connecting it to your own data (documents, mailbox, accounting), testing it on real questions until the answers hold, and keeping it running when a provider retires a model or changes a price. If the assistant talks to your customers, check the EU AI Act's transparency duties for chatbots; the EU AI Act checker tells you what applies, and do I need an AI policy covers your own team.

• Pick a model per task, not one for everything. A light model sorts e-mail, a workhorse writes the quote.
• Expect change. Two of the thirteen prices in this calculator change within three months. Software that treats the model as a setting switches in an afternoon.
• Measure what it returns. Count the hours a task took before you automated it. That is where the gain is, not in a cent per conversation.