What a context window costs you

A context window is advertised the way a fuel tank is advertised, in units of capacity, and billed the way fuel is billed, in units consumed. One million tokens is the headline figure on three vendors' pages this month. It is a ceiling, not an allowance. Nothing inside it is included, and on two of the three pages the price per unit changes once a request climbs past a line drawn inside the window. We read the Claude API pricing page, the OpenAI API pricing page and the Gemini API pricing page on October 14, 2026. This essay is about the three separate ways that headline number turns into money, and about a fourth that moves the number underneath you.

The headline number is a ceiling you pay to approach

Anthropic's pricing documentation is unusually direct about this. It states that "Claude 4.6 and later models (except Claude Haiku 5.5) and Claude Mythos Preview include the full 1M token context window at standard pricing", and then adds the clarifying parenthesis: "A 900k-token request is billed at the same per-token rate as a 9k-token request."

That is a generous policy and it is also the misunderstanding in one line. The rate holds. The bill does not.

At the Claude Sonnet 5.5 input price of $2 per million tokens listed on the same page, a 9,000-token prompt costs about 1.8 cents of input and a 900,000-token prompt costs about $1.80. That is our arithmetic from the published rate.

One hundred times the context is one hundred times the input bill. The window is also refilled on every turn of a conversation unless something is cached, so the multiplication repeats. The capacity is free of charge. Filling it is the product.

Two of the three vendors charge more at the top of the window than at the bottom

OpenAI's pricing page splits every flagship row into Short context and Long context columns and defines the boundary in a note beneath the tables: "Short context: 272K input tokens or fewer. Long context: more than 272K input tokens."

The step is a doubling of input. On that page, read October 14, 2026, gpt-6-astra is listed at $10.00 per million input tokens short and $20.00 long, with output at $50.00 and $75.00. gpt-6.1-sol is $2.00 and $4.00 input, $10.00 and $15.00 output. gpt-6-luna is $0.10 and $0.20 input, $0.50 and $0.75 output.

Google draws the same line lower. Gemini 3.1 Pro Preview is listed at $2.00 per million input tokens for prompts up to 200k and $4.00 for prompts over 200k, with output at $12.00 and $18.00.

Notice which number moves. On both pages the output price rises when the input is long, even though the answer is no longer than it would have been. You are not billed only for what you asked for. You are billed at a rate set by how much you handed over.

The line is a cliff, and a cache hit does not get you under it

Anthropic's one exception is the clearest version of the mechanism. Claude Haiku 5.5 is listed at $0.10 per million input tokens "for prompts up to 100,000 tokens" and $0.50 "for prompts over 100,000 tokens". Cache reads move from $0.01 to $0.05 at the same line.

Five times the price, at a threshold, not along a gradient. A request of 100,001 tokens is priced on a different scale from a request of 99,999, and the documentation says each request is priced on its own.

The same page then closes the obvious escape route. "A request's prompt length counts all of its input tokens, including cache reads and cache writes," it says, and a request over the threshold "pays the higher prices even when part of its prompt is a cache hit".

So caching lowers the rate on repeated content without shortening the prompt for pricing purposes. The architecture decision and the billing decision are the same decision, which is not how the window is described when it is being sold.

On one page, holding context is rent rather than purchase

Google's page adds a line the other two do not have. Gemini 3.1 Pro Preview context caching is listed at $0.20 per million tokens for prompts up to 200k and $0.40 over, plus a separate charge of "$4.50 / 1,000,000 tokens per hour (storage price)".

Per hour. A 500,000-token cache held across an eight-hour working day is $18 of storage by our arithmetic from that rate, before a single request reads it.

Gemini 3.8 Flash's storage line is $0.50 per million tokens per hour "through December 31, 2026" and "$1.00 / 1,000,000 tokens per hour (storage price) starting January 1, 2027". The doubling is already printed on the page we read.

Anthropic prices caching as multipliers on the base input rate instead: a five-minute cache write at 1.25x, a one-hour write at 2x, and a read at 0.1x, falling to 0.05x on Claude Opus 5.5 and Claude Sonnet 5.5.

There is no clock in that scheme, but the write is the commitment. The documentation notes that caching pays off after one read at the five-minute duration and two reads at the one-hour duration.

The token is not a fixed unit of text

The quietest line is in the notes under Anthropic's model table: "Claude 4.7 and later models and Claude Mythos Preview use a newer tokenizer that contributes to their improved performance on a wide range of tasks. This tokenizer produces approximately 30% more tokens for the same text."

Read that against a per-token price and a per-token window. The same document becomes roughly 30 per cent more expensive to send and occupies roughly 30 per cent more of a window whose headline number did not change. The page says the exact increase depends on content and workload shape.

Nothing is concealed here. It is also why a cost model built on last quarter's token counts can be wrong by a third without anyone editing a prompt.

Budget the refill, not the capacity

The useful budget line is not the window size. It is the four things the window does to a bill:

• The refill. Tokens per turn multiplied by turns per day, because an unmanaged conversation resends its own history every time.
• The threshold you will cross. Find the vendor's line - more than 272K input tokens on the OpenAI page, over 200k on the Gemini page, over 100,000 for Claude Haiku 5.5 - and decide whether your work sits above or below it, because the price changes at the line rather than gradually.
• The cache as rent or as multiplier. Google prices held context by the hour; Anthropic prices the write at 1.25x or 2x. Both reward long-lived context and punish context nobody reads.
• A tokenizer allowance. When a vendor changes tokenizers, the same text costs more. Anthropic has published approximately 30 per cent for its newer models.
• The overhead you did not write. Anthropic's documentation lists 286 tokens of tool-use system prompt for Claude Opus 5.5, charged before your own instructions arrive.

The arithmetic for two, ten and twenty people sits in our primer on AI assistant pricing for small teams, and the subscription side of the same question is in what a seat means when the worker is software. The two vendors' consumer-facing plans are set side by side in our ChatGPT and Claude comparison.

Why the word window survives anyway

The metaphor persists because it answers the question buyers actually ask, which is whether the thing will fit. A window implies a frame, and a frame implies a yes or no.

What the metaphor cannot carry is that the frame is priced by the square inch, that the price changes partway up, and that the inch itself was redefined when the tokenizer changed. A window you pay to look through is a different object from a window.

This is the same trick we described in unlimited is a pricing word: a capacity word doing commercial work. The difference is that unlimited hides a limit, while a context window advertises one and hides the meter.

What follows in practice

Price a workload, not a window. Take a real day of work, count the tokens it sends, multiply by the rate for the side of the threshold you land on, and add the cache line.

Then check the figure again when the vendor ships a new model family, because the tokenizer note is where the surprise lives. Per-tool detail sits in our subject pages for ChatGPT, Claude, Cursor and GitHub Copilot.

And keep the older argument in view: as we put it in the tool is not the workflow, the expensive context is usually the useful one.

A window full of your own documents is what the tool is for. Knowing it is metered is not a reason to leave it empty. It is a reason to stop calling the number an allowance.

Sources, read October 14, 2026: docs.claude.com pricing documentation; platform.openai.com API pricing; ai.google.dev Gemini API pricing.