models

Moonshot ships Kimi K3.1 Pro at $2/$10 per million tokens, with weights promised for November

Moonshot released Kimi K3.1 Pro on October 23: a 2.8T-parameter mixture-of-experts model (104B active) with a 1M-token context window. It's on the Kimi API today for $2 per million input tokens and $10 per million output, cheaper than K3's $3/$15. The open weights aren't out yet.

Written by Mara at Sig · Oct 24, 2026

K3.1 Pro is a post-trained version of K3, and the API model IDs are kimi-k3.1-pro and kimi-k3.1-pro-highspeed. The endpoint is OpenAI-compatible. Highspeed runs 2 to 3x faster by Moonshot's count and costs double: $4 input, $20 output. Cached input is $0.20 per million on both, and pricing is flat across the whole 1M context.

A few quirks from the quickstart docs: thinking is always on, reasoning_effort takes low, high or max (max is the default), and temperature and top_p are fixed, so leave them out of your requests. You also have to send back the full assistant message on every turn, and image input won't take public URLs (use base64 or an uploaded file ID).

On benchmarks, Moonshot's own numbers (self-reported, mostly run in the Claude Code harness) put it at 65.7% on Terminal-Bench 4.0, just behind Claude Opus 5.5 at 66.4%. It trails badly on science terminal tasks: 38.1 against 68.1 for GPT-6 Astra in Kimi's own chart.

Why it matters

If you route coding or long-context agent work to a frontier closed model, this is a cheaper option with a 1M window that you can call through the same OpenAI client. Default max reasoning can eat output tokens, though, so the per-token price isn't the whole bill.

Try this

Point your OpenAI client at the Kimi API, set the model to kimi-k3.1-pro, and run the same task twice: once with reasoning_effort low and once at the default max. Compare the output token counts in the usage field before you pick a default. You need a $1 top-up before the API takes requests.

Sources

Back to all news Get Sig Weekly every Monday