Kimi K3.1 Pro: what it is, pricing and how to use it

Kimi K3.1 Pro is Moonshot AI's new flagship model, released on October 23 with an API price of $2 per million input tokens and $10 per million output tokens. You can use it right now in the Kimi apps, Kimi Code and the API, and Moonshot says the full open weights will be published by November 4.

Moonshot announced it on X as "our most capable 3T-class open-weight model." The specs in that post: a mixture-of-experts design with 2.8 trillion total parameters and 104 billion active per token, a 1 million token context window, and image input. It's a post-trained version of Kimi K3, the July model that topped the frontend code arena, with substantially more post-training compute, much of it large-scale reinforcement learning.

What the independent numbers say

Launch-day benchmarks usually come from the company selling the model. This time an outside lab posted results within half an hour. @ArtificialAnlys says K3.1 Pro "launches at 4th on the Artificial Analysis Intelligence Index scoring 55." That's 11 points above K3. The only models ahead of it are Claude Opus 5.5 at 58 and Claude Sonnet 5.5 and Muse Wonder Full, both at 56.

Other numbers from the same post:

• Terminal-Bench 4.0: 67% in their own run, which they say is the top score of any model. Moonshot's launch post reported 65.7%, just under Opus 5.5.
• AutomationBench-AA, which tests agentic SaaS workflows, rose from 58% to 73%.
• Hallucination rate on their AA-Omniscience test fell from 53% to 44%.
• Cost to run their full index came to about $1.98 per task, almost the same as K3. Prices dropped 33%, but the model writes about 14% more output tokens.

If you're budgeting, that last line is the one to read twice. @natolambert caught the same thing: per token it's cheaper than K3, but it answers at greater length, so your bill may not move much. He also wrote that the scores would make it "clearly the strongest current open weight model," provided the weights actually ship on schedule.

Pricing at a glance

Artificial Analysis confirmed the standard-model prices. The Highspeed rates come from an OrcaRouter launch writeup that cites Moonshot's pricing page. Kimi K3 was $3 and $15.

How to use Kimi K3.1 Pro

Moonshot's launch post lists four ways in:

1. Chat: update the Kimi app or open kimi.ai. The new 3D Make feature is there too: it builds a 3D model you can edit in the browser and download ready to print.
2. Kimi Work: the desktop app, version 3.1.0 or later, on Windows or Apple silicon Macs.
3. Kimi Code: type /model in the CLI and pick K3.1 Pro. The Kimi Code docs say it requires the Plus plan or higher, the full 1M context requires Pro, and Highspeed requires Pro as well.
4. API: call kimi-k3.1-pro from the Kimi API Platform. Zero Data Retention is available.

A few gotchas from the docs. The model only works with thinking on, so a request that disables thinking returns HTTP 400. Reasoning effort accepts low, high or max, and max is the default. A typo in the model ID quietly sends you to K2.8 Preview with no thinking, so copy the ID exactly. Changing models or effort mid-session also clears the context cache, so start a new session when you switch.

For businesses using Kimi Enterprise, the model ships turned off and an admin has to enable it.

The cyber trade-off

Moonshot was unusually open about one choice. Before the open-weight release, it removed cyberattack-related data from training. On its internal exploit benchmark, a checkpoint trained without that filtering solved 90.5% of tasks and the released model solved 81%. K3 solved 33.3%. Moonshot says its goal is to stay behind the open-weight cyber frontier and raise that capability gradually. Once weights are downloadable they can't be recalled, so expect security teams to test that claim hard after November 4. By Moonshot's own count, DeepSeek's open V4.1 Flash scores 85.7% on the same test, so K3.1 Pro is not the most capable open model here.

What this means for a small business

A model that scores near Claude Opus 5.5 on agent tasks, at a fraction of the price and with downloadable weights, keeps pushing AI costs down. Open models are already handling a growing share of real workloads, as we noted in our look at open-weight token share. For most owners the practical move is to test one real workflow on two models and compare quality and the actual bill, rather than trusting launch charts.

If you'd like help choosing which of your processes fit a cheaper model and which still need a premium one, our free process audit starts there.