# Tokens and context windows: what actually drives your AI bill

- **Published:** 2025-12-04
- **Updated:** 2026-10-12
- **Category:** Economics & Pricing
- **Author:** Sinthetix
- **URL:** https://sinthetix.com/blog/tokens-context-windows-ai-bill

> Model APIs bill by the token, and long prompts repeated many times add up. A short guide to reading the price list and one way to cut it.

Language model APIs charge by the token, a chunk of text roughly the size of a short word or word fragment. Two numbers on a price list matter most: the price per million input tokens (what you send) and per million output tokens (what the model writes back). They are not the same price.

**Key takeaways**

- Output tokens cost several times more than input tokens on current frontier models.
- A context window is the most text a model can consider in one request; every request re-sends whatever you put in it.
- Repeating a large document in every call is the usual hidden cost, and prompt caching targets exactly that.

## Reading the price list

Anthropic prices Claude Opus 4.5 at $5 per million input tokens and $25 per million output tokens [1]. Google lists Gemini 3 Pro, in preview, at $2 and $12 for prompts of 200,000 tokens or less [4]. In both cases output costs five to six times as much as input, so a task that generates long answers costs more than one that mostly reads.

The context window is the ceiling on what fits in one request. For scale, Anthropic describes 200,000 tokens as roughly 500 pages of text [3]. Larger windows are useful, but they are not free: you pay for every token you send, on every call.

## The hidden multiplier

Consider a hypothetical support assistant that includes a 100,000-token policy manual in every question. At a price of $3 per million input tokens, that is $0.30 of input per question before the model writes a word, or about $300 across 1,000 questions. The dollar figures are derived: our own arithmetic on the $3 per million input rate Anthropic published for Claude 3.5 Sonnet in its caching announcement [2].

Prompt caching changes this. Anthropic states that reading cached input costs 10% of the base input price, with writing to the cache costing 25% more than base, and reports cost reductions of up to 90% on long prompts [2]. On the same example, repeated reads would cost roughly $30 plus a one-time cache write of about $0.38, again derived from the published rates.

| Approach (1,000 questions, 100K-token manual) | Approx. input cost (derived) |
| --- | --- |
| Resend manual every time | $300 |
| Cache the manual, read it each time | about $30 |

The saving depends on request volume and on each vendor's cache terms, so check current terms before budgeting.

## What to do about it

- Ask which tokens dominate your bill: long inputs repeated across calls, or long outputs.
- Send only what the task needs. Retrieval that supplies a few relevant passages is often cheaper than a whole manual.
- Use caching where the same large block precedes many requests.

The practical implication: estimate cost per task, not per token, by multiplying realistic prompt and answer lengths by expected volume before choosing a model.

## Sources

1. [Introducing Claude Opus 4.5](https://www.anthropic.com/news/claude-opus-4-5), Anthropic, 2025-11-24
2. [Prompt caching with Claude](https://claude.com/blog/prompt-caching), Anthropic, 2025-08-14
3. [Introducing Contextual Retrieval](https://www.anthropic.com/news/contextual-retrieval), Anthropic, 2024-09-19
4. [Gemini 3 for developers](https://blog.google/technology/developers/gemini-3-developers/), Google, 2025-11-18

## How Sinthetix helps

An engineering studio for AI agents, automation, and internal tooling. We build the systems that make operators scale.

[Our services](https://sinthetix.com/services) | [Talk to us](https://sinthetix.com/contact)

Sinthetix is an AI engineering studio. We ship agents, automation, and internal tools for operators who need results, not reports. [About Sinthetix](https://sinthetix.com/about)