← back to the archiveCover illustration for “Token pricing: how AI model APIs charge you”
TERM8mo ago·Published · Updated ·by Andy Padia

Token pricing: how AI model APIs charge you

In short: Token pricing bills AI model API use per million input and output tokens, with modifiers for caching, batch jobs, long prompts and tools.

Token pricing is the way most AI model APIs charge you: one rate per million input tokens, which you send, and a usually higher rate per million output tokens, which the model generates. The OpenAI, Anthropic and Google price pages, read in November 2026, also list cached-input rates and cheaper batch rates.

Token pricing in plain words

Think of a taxi meter that charges per kilometre, with a higher rate for the stretch where the model does the talking.

How it works

Every response reports its input and output token counts. Your bill is input tokens times the input rate plus output tokens times the output rate, then adjusted by modifiers. On the three pages above, as read in November 2026, the modifiers look like this.

Cached input is billed below the base input rate; Anthropic lists cache reads at 0.1 times base input on most models. Batch jobs cost about half, and Anthropic and Google both state a 50 percent reduction. Thinking tokens count as output on Google's page, which prices output including thinking tokens. Long prompts can move you to a higher tier: some Gemini models charge more for prompts above 200,000 tokens. Tools are metered separately, and OpenAI and Anthropic each list web search at $10 per 1,000 calls on top of tokens.

Subscriptions for chat and coding products bundle usage differently from APIs.

Example

As an illustration of the arithmetic, with made-up rates: 2,000 input tokens at ₹250 per million and 500 output tokens at ₹1,000 per million cost ₹0.50 plus ₹0.50, so ₹1 per request. A million such requests a month is ₹10 lakh, before any caching or batch discount. Always check the provider's current price page.

Where you'll meet it

  • Provider pricing pages and usage dashboards
  • Invoices from cloud model hubs
  • Budget planning for AI features
  • Comparisons between API and self-hosted costs

What it changes for you

For developers, output length is often the biggest lever, because output is the dearer side of the rate card: concise responses and limited reasoning effort save money. For finance, my view is that the rate card is only half the bill. Agent loops multiply token volume, so track volume as closely as rates. For product teams pricing AI features, token costs must be translated into cost per task and then into a margin.

Token pricing vs subscription pricing

Token pricing charges per use. Subscriptions charge a flat fee with usage limits, which suits individuals but hides the true per-task cost.

Common mistake

My view: comparing models on the input price only. Many workloads are dominated by output and reasoning tokens. Model your actual input-to-output ratio before choosing.

Token · Cost per task · Prompt caching · Batch inference · Reasoning model

From the archive: How much does an LLM feature cost in production? · Task cost is a harness property, not a model price

Compare models on your real input-to-output ratio, never on the input price alone.

#field-guide#ai-economics#pricing#cost#tokens

Sources

  1. developers.openai.com/api/docs/pricing
  2. platform.claude.com/docs/en/about-claude/pricing
  3. aistudio.google.com/docs/pricing
← older drop
Prompt caching: paying less for repeated context
newer drop →
Tokens per second: how fast an AI model writes

related drops

explore all 557 drops →
← back to the archiveday 175