Moonshot's Kimi K3.1 Pro undercuts K3 by a third, weights due by Nov. 4
It costs $2/$10 per million tokens and ranks fourth on Artificial Analysis' index, but the promised open weights and their license aren't out yet.
Moonshot AI released Kimi K3.1 Pro on Oct. 23, a more heavily post-trained version of its Kimi K3 model that costs a third less through the Kimi API and is promised as open weights by Nov. 4.
The model is "a post-trained version of Kimi K3, with substantially larger post-training compute," Moonshot said in its launch post. It keeps a 1-million-token context window. Moonshot calls it "3T-class," and Artificial Analysis lists 2.8 trillion total and 104 billion active parameters.
A second model, K3.1 Pro Highspeed, costs twice as much for input and output, and Moonshot said it "delivers 2-3× the speed of K3.1 Pro." EngineerPal found no independent speed test of Highspeed.
Several things aren't known yet. Moonshot hasn't published a license for the weights, and Artificial Analysis lists the model as "Proprietary" until they ship. It's also unclear whether the open weights will match the API model, since Moonshot says it removed cyber-related training data "for the open-weight release."
Kimi K3.1 Pro price and how to use it
The Kimi API pricing page lists these rates in US dollars per million tokens, excluding tax. K3.1 Pro is 33% cheaper than K3 on every rate. Highspeed costs 33% more than K3 for input and output.
| Model ID | Input | Output | Cache hit |
|---|---|---|---|
kimi-k3.1-pro |
$2.00 | $10.00 | $0.20 |
kimi-k3.1-pro-highspeed |
$4.00 | $20.00 | $0.20 |
kimi-k3 (previous) |
$3.00 | $15.00 | $0.30 |
The model always reasons, and reasoning tokens are billed as output. Artificial Analysis measured 55,000 output tokens per task on its index, 14% more than K3. Pricing doesn't change with context length, according to the quickstart.
K3.1 Pro is available on Kimi.ai, the Kimi apps for iOS, Android and HarmonyOS, Kimi Work 3.1.0 or later, and Kimi Code, where /model switches to it. Kimi Enterprise access "is off by default at launch," and admins must turn it on.
The API is OpenAI-SDK compatible at https://api.moonshot.ai/v1, and it unlocks K3.1 Pro after a top-up of at least $1. The reasoning_effort setting accepts low, high or max, with max as the default. This request comes from Moonshot's quickstart:
curl https://api.moonshot.ai/v1/chat/completions \
--header "Authorization: Bearer $MOONSHOT_API_KEY" \
--header "Content-Type: application/json" \
--data '{"model": "kimi-k3.1-pro", "messages": [{"role": "user", "content": "Introduce Kimi K3.1 Pro in one sentence."}]}'
The docs say to send back the complete assistant message in multi-turn and tool calls, not just its content. Moonshot's web search tool "is being updated and is not recommended" for production. Vercel AI Gateway also offers the model as moonshotai/kimi-k3.1-pro.
The launch also starts the wind-down of Kimi K2. K2-series models close to new users on Oct. 31, and Moonshot plans to retire them by the end of December, according to its pricing page.
Kimi K3.1 Pro benchmarks and the open-weights date
In independent testing, Artificial Analysis ranks K3.1 Pro fourth of 215 models on its Intelligence Index, with 55. Claude Opus 5.5 (58), Claude Sonnet 5.5 (56) and Muse Wonder Full (56) score higher. Kimi K3 scored 44.
Terminal-Bench 4.0 results depend on who ran the test. Moonshot's own table gives K3.1 Pro 65.7%, just behind Opus 5.5 at 66.4% and ahead of GPT-6 Astra at 58.2%. Artificial Analysis' run gave it 67%, "the highest score of any model," with Opus 5.5 at 60%.
Moonshot's other vendor-reported results include 92.7% on BenchCAD with tools, behind Opus 5.5 and GPT-6 Astra. Several of its gains are on internal benchmarks, including a hallucination test that Moonshot says doesn't reflect everyday error rates. Plotly said it scored 42 of 43 on the company's own data-analytics exam.
Artificial Analysis' own knowledge test found a hallucination rate of 44%, down from K3's 53%, with accuracy rising from 48% to 51%. It also called the model "very verbose" and measured output at about 73 tokens per second, below its average of 82.
"Full model weights will be released by Nov 4, 2026," Moonshot said. Artificial Analysis says K3.1 Pro would then be the top open-weights model on its index.
