Kimi K3.1 Pro context window and output limit

Kimi K3.1 Pro reads up to a million tokens and, as Kimi K3.1 Pro Outlaw, writes up to 943,718 of them in a single reply: the longest output of any model Wild West sells.

The two numbers

The window is shared: the prompt and the reply have to fit in it together. A 900,000-token prompt leaves about 148,576 tokens for the answer, whatever max_tokens asks for. Reasoning is written into the same output budget as the answer, so a hard question uses part of it before the answer starts.

Against the rest of the line

Every model on the line reads about a million tokens. What sets Kimi K3.1 Pro Outlaw apart is how much it can write back in one go, which matters for long reports, whole generated files and reasoning-heavy answers.

What filling the window costs

Reading a full million tokens fresh costs $7.50 at $7.50 per million. Sent again, the repeated part bills as cached input at $0.80 per million, so a second read of the same million costs $0.8000. That is why an agent that re-sends its history every turn costs far less than the headline input price suggests.

Long replies and spend caps

A large max_tokens does not mean a large bill. Wild West prices the worst case before the model runs, holds it against the key's cap, and lowers max_tokens to what the hold covers, then charges only what was written. See max_tokens clamping and per-key spend caps.

FAQ

Related