What Is Context Length? (And What Actually Fills It)

Every model card carries a number like "200K context" or "1M context". It looks like a capacity spec, the way disk space is. It is closer to a budget you spend on every single request — and most of what spends it is not the thing you typed.

Figures last checked: August 26, 2026.

What context length is

A model has no memory. Each time you call it, you send the entire conversation again — system prompt, every previous turn, any attached files — and it reads all of it before writing a word.

Context length is the maximum size of that bundle, measured in tokens. A token is a small piece of text, roughly three-quarters of a word in English.

So it is not storage. It is the size of the desk the model can lay your documents out on, refreshed from scratch every request.

The real limits today

Two things are worth noting.

A million tokens is roughly 750,000 words — several long novels. Almost no application needs that, and the ones that do usually have a retrieval problem rather than a context problem.

1,048,576 is 2²⁰, not a marketing round number. Where a provider quotes exactly 1,000,000 they have rounded; where they quote 1,048,576 they are reporting the real buffer size.

Token prices and context lengths for every text model are on the chat models page.

What actually fills it

This is the part model cards do not tell you. Anthropic's documentation publishes exact token counts for a few minimal requests, and they are instructive:

One tool definition costs roughly 390 tokens. An agent with twenty tools spends about 8,000 tokens of its window before the user has typed anything — and pays for them on every request.

An image is about a thousand tokens. Twenty screenshots is 20,000.

And the conversation itself grows. Turn 10 carries turns 1 through 9. A chat that felt small at the start is the largest thing in the request by the end.

So a "200K context" model in an agent with fifteen tools, a long system prompt and a few screenshots is not starting at 200K. It is starting somewhere well below it, and shrinking with every turn.

The trap: the same text is not the same number of tokens

Context length is measured in tokens, and tokens are not a stable unit across models.

Anthropic's own documentation states that Claude 4.7 and later use a newer tokenizer where "the same input text produces approximately 30 percent more tokens than on earlier models."

Which means: a prompt measured at 150,000 tokens on an older model can be about 195,000 on a newer one — the same words, thirty percent more of the window, and thirty percent more cost. If you sized a document-processing pipeline against one model and then switched, re-measure before assuming it still fits.

Providers give you a way to check. Anthropic's token-counting endpoint is free to call. Count against the exact model you plan to run.

What happens when you run out

Not a graceful degradation. Depending on the provider you get an error, a truncated input, or silently dropped history — and the third is the dangerous one, because the model answers confidently from a conversation it can no longer fully see.

Three ways out, in order of how much work they are:

Trim. Drop or summarise the oldest turns. Cheapest to build, and usually enough for chat.

Cache. If the large part of your context is a fixed prefix — a long system prompt, a reference document — prompt caching lets the provider store it and charge a fraction to re-read it. This does not enlarge the window, but it removes most of the cost of filling it.

Retrieve. Stop sending the whole corpus and send only the relevant parts. More engineering, but it is the only approach that does not get worse as your data grows.

Reaching for a bigger context window is the fourth option, and usually the most expensive one — because you pay for every token in that window on every request.

Does a bigger window cost more?

Not per token. A 1M-context model is not billed at a higher rate than a 200K one for the same request; Claude Sonnet 5 is $2.00 per million input tokens whether you send five hundred tokens or five hundred thousand.

But a bigger window makes it easy to spend more, because it removes the error that would have stopped you. Filling a million-token window at $2.00 per million costs $2.00 per request — and if that request is a chat turn happening fifty thousand times a month, the arithmetic gets serious quickly.

The discipline is the same either way: send the least you can, not the most you are allowed to.

The short version

• Context length is the maximum size of one request, in tokens, not a memory.
• Today's limits run from about 33K to about 1M; most production work fits comfortably at the low end.
• Tool definitions (~390 tokens each), images (~1,000), PDFs (~2,000) and re-sent history fill it faster than your prompts do.
• The same text is a different number of tokens on different models — recheck when you switch.
• A bigger window does not cost more per token. It just makes it easier to spend more of them.

Explore more on ElliSekiz

Compare context lengths and token prices side by side. The chat models page lists every text model with its input and output rates.

Then measure your own prompt. Every model page has a playground — run a real request and read the actual token counts back before you size anything.

Sources

• OpenRouter models API, https://openrouter.ai/api/v1/models (context lengths, read 2026-08-26)
• Anthropic — Token counting (documented example counts, tokenizer change, free counting endpoint)