What is the difference between AI tokens and credits?

A token is the unit an AI model reads and writes, and the unit the labs bill in. A credit is a prepaid balance you buy in dollars and spend down on tokens, sometimes at the lab’s token price and sometimes through a conversion table of the lab’s own. Tokens are the cost; credits are one way of paying for it.

What a token is

OpenAI’s token guide describes a token as “a character, part of a word, a whole word, or punctuation,” and gives three rules of thumb for English: one token is about 4 characters, one token is about three quarters of a word, and 100 tokens is about 75 words. Anthropic’s glossary puts a Claude token at roughly 3.5 English characters, and its pricing FAQ uses the same 4 characters or 0.75 words. Code, numbers and other languages tokenize less efficiently, so the same page of text can cost more tokens in Japanese or JSON than in English prose.

Tokenizers also differ between models. Anthropic’s pricing page says Claude 4.7 and later models use a tokenizer that “produces approximately 30 percent more tokens for the same text” than earlier Claude models, so a prompt that counted 10,000 tokens on Sonnet 4.6 counts around 13,000 on Fable 5.1 before any price is applied.

Every request has two token counts. Input tokens are everything you send: your message, the system prompt, attached files and the whole conversation so far, which is resent on every turn. Output tokens are what the model writes back. Reasoning models add a third kind, and the labs bill it as output. OpenAI’s guide says reasoning tokens “are not visible as answer text, but they count toward output usage and are billed as output tokens.” Anthropic’s thinking documentation says thinking tokens “are billed as output tokens, even when the thinking text isn’t returned to you.” Google’s Gemini pricing page lists its output price as “including thinking tokens.”

How labs price per million tokens

Each lab publishes a price per million tokens, with input cheaper than output. On October 1, 2026 the Claude API pricing page lists Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens, Claude Opus 5.5 at $4 and $20, and Claude Fable 5.1 at $10 and $50. OpenAI’s API pricing page lists GPT-6.1 Sol at $2 and $10, GPT-6 Astra at $10 and $50, and GPT-6 Luna at $0.10 and $0.50. Cached input, where the provider recognizes a prompt prefix it has processed recently, costs a tenth of the input price or less at both labs, and batch jobs that can wait hours get 50 percent off.

The per-million figure is the whole price. There is no subscription behind an API key; the bill is tokens multiplied by rate, which is why API spend on an ordinary chat comes to fractions of a cent per message.

How credits map to tokens

Consumer plans hide tokens behind a flat fee and a usage allowance. When the allowance runs out, both OpenAI and Anthropic sell credits, and the two map to tokens differently.

Claude usage credits are dollars. Anthropic’s usage credits article says they are “billed at standard API rates,” so a Pro subscriber who buys $20 of credits gets the tokens that $20 buys on the API. Usage bundles discount them: $50 of usage for $45, $250 for $200 and $1,000 for $700.

ChatGPT credits are a unit of their own. OpenAI’s pricing documentation publishes a rate card in credits per million tokens: GPT-6 Luna costs 2.5 credits per million input tokens and 12.5 per million output, GPT-6.1 Sol 50 and 250, and GPT-6 Astra 250 and 1,250, with fast mode at twice the standard rate. The credits article says credits work in Codex, ChatGPT Work and the Office add-ins rather than in ordinary chat, that a typical Excel task uses 5 to 20 credits, and that they are valid 12 months from purchase. Neither page prints a dollar price per credit; OpenAI says purchase amounts and prices vary by account, region and plan, so the conversion to dollars is on your own receipt.

API credits are a third thing again: prepaid dollars for a developer account, with a $5 minimum at OpenAI, that expire a year after purchase there and at Anthropic. OpenAI’s credits article is explicit that ChatGPT credits “are not API credits.”

Why the same dollar buys different token counts

Because the per-token price differs by model, by direction and by tokenizer. One dollar buys 500,000 input tokens on Claude Sonnet 5.5 and 100,000 on Claude Fable 5.1; the same dollar buys 10 million input tokens on GPT-6 Luna. Output costs five times input on all three. A cheaper model can still cost more if it splits your text into more tokens or reasons at length. OpenAI’s token guide says “a lower price per million tokens does not necessarily produce a lower total cost.”

A worked example

Take one chat turn: you paste a 900-word email thread and ask for a reply, about 1,200 input tokens, and get a 300-word draft back, about 400 output tokens.

Model	Input cost	Output cost	Turn cost	Turns per $1
Claude Sonnet 5.5 ($2 / $10)	$0.0024	$0.0040	$0.0064	156
Claude Fable 5.1 ($10 / $50)	$0.0120	$0.0200	$0.0320	31
GPT-6 Luna ($0.10 / $0.50)	$0.00012	$0.00020	$0.00032	3,125

In ChatGPT credits for Codex, the same turn on GPT-6.1 Sol is 1,200 tokens at 50 credits per million plus 400 tokens at 250 credits per million, or 0.16 credits. On Claude usage credits it is the Sonnet or Fable line above, in dollars, because Anthropic charges API rates. If the turn runs on a reasoning model with thinking turned up, add the hidden tokens at the output rate before trusting either figure.

Brass sells prepaid packs of $10 to $250 that work across Claude, GPT, Gemini, Grok and DeepSeek, billed per message at the model’s list price plus 25 percent, with the token count on each receipt.

For what a fixed budget buys on each current model, see How many tokens does $25 buy on each AI model?

Sources

Related guides

How many tokens does $25 buy on each AI model?
At October 2026 list prices, $25 buys 1.25 million tokens on Claude Fable 5.1 or GPT-6 Astra and 125 million on GPT-6 Luna. The table, with message counts.
Do OpenAI API credits expire?
Yes. OpenAI’s prepaid API credits expire one year after purchase, are non-refundable and cannot be extended. What that covers and how to check your balance.
Is there a pay-as-you-go version of ChatGPT?
Not for the chat app. ChatGPT sells $8, $20 and $100 to $500 monthly plans. Pay-per-use exists through the API and add-on credits, and most messages cost cents.