Context window
A context window is the maximum number of tokens a model can hold in one request, counting the system prompt, the conversation history, any retrieved documents and the reply being generated together.
The window is a hard ceiling on a single request, not a memory. Nothing carries between requests unless the client sends it again, which is why a chat client resends the whole conversation on every turn and why the request grows steadily as the conversation does.
Exceeding the window is an error, not a silent truncation, on any correctly implemented API. That is the behaviour you want: a request quietly trimmed to fit produces an answer based on material the model never saw, and the failure surfaces later as a wrong answer rather than immediately as a rejected request.
Usable context and advertised context are not the same thing. Retrieval accuracy over very long inputs degrades well before the stated limit on most models, so a window figure is a capacity claim rather than a quality claim. Filling one entirely is rarely the best use of it.
Window sizes in this catalogue are quoted in tokens rather than words, and vary by roughly a factor of four between the smallest and largest text models we serve.
What it costs you
Every token inside the window is billed on every request that carries it, so context is a recurring cost rather than a capability you buy once. A long system prompt is not paid for at the start of a conversation; it is paid for again on each turn. This is also why conversation cost grows faster than message count — with an unbounded history, the tokens sent across a conversation scale with the square of its length, which is the single most common reason a chat application costs several times what its author modelled.
Worked from today’s rates
Kimi K3 has a 512K context window — around 393,000 words of English at once. Filling it costs $0.9699 in input on every request that carries it, at $1.85/M.
Related terms
- TokenA token is the unit of text a language model reads and writes — roughly four characters of English, or about three quarters of a word — and it is also the unit that every language model API bills by.
- Prompt cachingPrompt caching is the reuse of an already-processed prompt prefix, so that repeated leading tokens — a system prompt, a tool schema, a fixed document — are billed at a reduced cached-input rate instead of the full input rate.
- Input vs output tokensInput tokens are the text you send a model and output tokens are the text it writes back; they are billed at separate rates, and the output rate is almost always several times the input rate.
Go deeper
Every rate quoted above is published in full on the pricing page, alongside the model publisher’s own official rate.