Kimi K3.1 Pro: $2 input, weights promised by Nov 4

Kimi K3.1 Pro showed up on the Kimi API on October 23, 2026. The catch for anyone who wants to run it locally: the weights aren't out yet. Moonshot says full model weights will be released by November 4, 2026, according to Simon Willison · LLM notes.

So right now you can only reach it through the API. If your plan was to download it and run it on your own hardware, that date is the thing to watch.

The dollar pricing is $2.00 per million input tokens and $10 per million output tokens. That puts it at the same level as GPT-6.1 Sol and at half the price of Claude Opus 5.5.

There's also a Highspeed version, which Moonshot says is 2-3x faster and costs twice as much. The exception is input token cache hits, which stay at the same price.

Moonshot's own benchmarks put Kimi at roughly the same level as Opus 5.5 and GPT-6 Astra. It beats them on a few tests and comes close on others. Those are the vendor's numbers, so treat them as a claim, not a result.

A separate score tells a similar story. Artificial Analysis scored Kimi at 55, ahead of Claude Fable 5.1, GPT-6 Astra and Gemini 4 Argon, all at 53. Only Claude Opus 5.5 sits higher, at 58.

Willison calls this the first time a soon-to-be open weight model has scored this highly against the frontier. That's his read, not a verified fact, and it's the kind of claim worth checking once the weights actually ship.

Willison ran it through the Moonshot API using the `llm openai endpoint` command added in LLM 0.32. The model name is `kimi-k3.1-pro`, and it supports three reasoning levels: low, medium and max.

He generated SVG pelicans at each level and compared them against fresh runs on Kimi K3. That's a quick sanity check, not a benchmark. I haven't run it myself, so I can't tell you how it handles your own workload.

Two things are still open. First, the weights: November 4 is a promise, not a release. Second, the speed claim on Highspeed is Moonshot's own, and doubling the price for 2-3x throughput is a trade-off you should weigh against your own workload.

My advice: test it on your real prompts at the standard tier first. Don't rebuild your stack around a model whose weights you can't download yet.