A context window is the amount of text a model can hold in mind at once. Everything counts towards it: your question, the documents you paste, the model’s own reply, and the whole conversation so far. When you exceed it, the earliest material falls out of view. The model does not tell you this happened. It simply stops knowing things it knew a moment ago.

Tokens are the unit. A token is roughly three quarters of an English word, so a million token window is very approximately 750,000 words, or about eight or nine full length novels. That sounds like more than anyone needs. In practice, people fill it faster than they expect, because a moderately sized codebase, a year of email, or a set of legal filings gets there quickly.

Where the numbers stand today

As of August 2026, roughly a million tokens has become the industry standard rather than a differentiator. OpenAI’s GPT-5.6 family carries 1,050,000 tokens across Sol, Terra, and Luna. Anthropic’s Claude Opus 5 and Sonnet 5 both carry 1,000,000. Google’s Gemini 3.1 Pro carries 1,048,576. DeepSeek-V4-Flash, Qwen3.8-Max, and Kimi K3 are all at or around a million as well.

The outlier is xAI’s Grok 4.5 at 500,000 tokens, which is half or less than the rest of the field. Mistral Medium 3.5 sits at 256,000, though that is a deliberate trade for a model designed to run on four GPUs rather than a data centre.

The part the specification sheet hides

Advertising a window and pricing it evenly are different things. OpenAI bills GPT-5.6 Sol at 5.00 dollars per million input tokens, but requests above 272,000 input tokens are billed at twice the input rate and 1.5 times the output rate. So the last three quarters of that headline window costs double. Gemini 3.1 Pro does something similar: 2.00 dollars per million input up to 200,000 tokens, then 4.00 dollars above it.

If you are budgeting for a system that routinely fills the window, model the cost at the higher tier. The advertised price applies to the part of the window most people never reach.

Caching is the lever that actually matters

Most long context work sends the same material repeatedly. A codebase, a policy manual, a product catalogue. Every provider now prices cached input far below fresh input, and the gap is enormous. OpenAI charges 0.50 dollars per million cached input tokens against 5.00 for fresh, a tenth. DeepSeek charges 0.0028 dollars per million on a cache hit against 0.14 on a miss, a fiftieth.

Structuring a prompt so the stable material comes first and the variable part comes last is the single largest cost saving available to most teams, and it is a formatting decision rather than a model decision.

A big window is not the same as using it well

Capacity is not comprehension. A model can accept a million tokens and still weight the beginning and end of that material more heavily than the middle, a behaviour researchers have documented across model families for years. Providers publish the window size because it is easy to measure. How reliably a model retrieves a specific fact from the middle of a very long document is harder to measure and less often published.

The practical test is not whether the model accepts your document. It is whether it can quote the right paragraph from the middle of it, and whether it says so when it cannot.

What to actually do

  • Do not choose a model on window size alone. Nearly all of them are at a million now. It has stopped being a meaningful point of difference.
  • Check the pricing tier boundary for the model you are using, and where your typical request lands relative to it.
  • Put stable material first so caching can do its job, and check your provider’s cached input rate.
  • Test retrieval from the middle of a long document before you build a workflow that depends on it.
  • Shorter is usually better. If a well chosen 20,000 token excerpt answers the question, it will be faster, cheaper, and more accurate than the full million.

The window is a ceiling, not a target.