How many tokens does $25 buy on each AI model?

Between 1.25 million and 125 million tokens, depending on the model. At the list prices published on October 1, 2026, $25 buys about 1.25 million blended tokens on Claude Fable 5.1 or GPT-6 Astra, 6.25 million on Claude Sonnet 5.5 or GPT-6.1 Sol, and 125 million on GPT-6 Luna. A typical chat turn of 1,200 input tokens and 400 output tokens therefore costs between a thirtieth of a cent and three cents, and $25 covers somewhere between 780 and 78,000 of them.

The table

Every lab charges less for input tokens (what you send) than for output tokens (what the model writes). Chat traffic skews toward input, because the whole conversation is resent on each turn, so the table uses a blend of three input tokens for every output token: blended price = (3 x input + output) / 4. The first token column is at the lab’s list price; the second is at list plus 25 percent. Figures are rounded to three significant figures.

Swipe the table sideways for more columns.

The list prices come from each lab’s own pricing page: Anthropic, OpenAI, Google, xAI and DeepSeek. Four caveats sit on those pages. Google charges $4 and $18 instead of $2 and $12 once a Gemini 3.1 Pro Preview prompt passes 200,000 tokens, xAI doubles Grok 4.7’s rates at 200,000 tokens, and OpenAI charges $20 and $75 on GPT-6 Astra past 272,000 input tokens. DeepSeek’s page lists $0.30 and $1.20, and half that outside its peak hours of 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays; the table uses the full rate, which the US companies that host the model charge at every hour. Today’s price for every model, per million words, is on the market board, read from these pages every hour. And Anthropic’s page notes that Claude 4.7 and later models tokenize the same text into about 30 percent more tokens than earlier Claude models, so a million Fable tokens holds less text than a million Sonnet 4.6 tokens did.

What a million tokens is

OpenAI’s token guide and Anthropic’s pricing FAQ both put an English token at about 4 characters or 0.75 words. A million tokens is therefore around 750,000 words. Call a page 500 words and that is 1,500 pages read or written; the $25 column for GPT-6 Luna is about 190,000 such pages, and for Claude Fable 5.1 about 1,900.

Messages are the more useful unit. Take a typical chat turn as 1,200 input tokens (your question plus the context the app resends) and 400 output tokens (a 300-word answer). That is 1,600 tokens at the same 3:1 ratio as the table, so dividing the table by 1,600 gives turns per $25.

Swipe the table sideways for more columns.

xAI also adds about 1,230 input tokens of its own to every Grok 4.7 message, 1,152 of them cached at $0.50 a million. Counting them, $25 buys about 4,500 Grok turns at list and 3,600 at list plus 25 percent.

Long conversations cost more per turn than this, because the input side grows with every message. A chat that has reached 20,000 tokens of history sends all of it again each time, so a turn on Claude Opus 5.5 at that depth costs about $0.088 at list and $25 lasts about 280 turns rather than 1,950. Starting a new chat for a new topic is the cheapest habit there is.

Reasoning models bill tokens you never see

The output figure in the table is the visible answer only. Reasoning models think before they answer, and OpenAI, Anthropic and Google all bill that thinking at the output rate. OpenAI’s reasoning guide says reasoning tokens “are not visible via the API” but “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,” and the Claude Code cost page notes that the default thinking budget can run to tens of thousands of tokens per request. Google’s pricing page lists Gemini output as “including thinking tokens.” A hard question that triggers 5,000 thinking tokens on Claude Opus 5.5 costs $0.10 before the answer starts, about eight ordinary turns’ worth. For budgeting, treat the turn counts above as ceilings for easy questions, and expect a third to a tenth of them on work that makes the model reason.

The second column of each table is what the same $25 buys at Brass, which sells prepaid packs of $10 to $250 for eight of these models, all but GPT-6 Luna, at list price plus 25 percent, with no expiry and a token count on every receipt.

For how tokens turn into the credits that ChatGPT and Claude sell, see What is the difference between AI tokens and credits?

Sources

1. Claude API pricing
2. OpenAI API pricing
3. Google: Gemini Developer API pricing
4. xAI: API pricing
5. DeepSeek: Models and pricing
6. OpenAI: Understanding and counting tokens
7. OpenAI: Reasoning models
8. Anthropic: Thinking
9. Claude Code: Manage costs effectively

Related guides

• What is the difference between AI tokens and credits?

  A token is what AI models bill in; a credit is a prepaid balance spent on tokens. How labs price per million tokens and how ChatGPT and Claude credits convert.

• 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.

• Will Claude usage limits increase in 2026?

  Anthropic’s published Pro and Max limits, the five-hour sessions and weekly caps, the September 14, 2026 Claude Code change, and what extra usage costs.