Kimi K3.1 Pro: near Opus, half the price.

Moonshot AI launched Kimi K3.1 Pro on Friday 2026.10.23. It is the first model in the K3.1 series, it ships alongside a faster K3.1 Pro Highspeed, and on Moonshot's own terminal-coding numbers it lands 0.7 points behind Claude Opus 5.5 at half the price. There is no plain "Kimi K3.1": the model is K3.1 Pro. The open weights are promised by 2026.11.04, and as of 2026.10 they aren't out.

Watch the 60-second version on Instagram.

Same engine, more training

K3.1 Pro is not a new architecture. It is Kimi K3's base with much more reinforcement-learning post-training on top. Moonshot calls it a "3T-class" model, with a 1M-token context (1,048,576 tokens) that also takes image and video input.

You'll see "2.8T total, 104B active" quoted for it. Those are Kimi K3's numbers, a mixture-of-experts model where only part of the network runs per token. Moonshot gives no separate figures for K3.1 Pro, so anything more precise than "3T-class" is inference.

The price

Per million tokens. K3.1 Pro is cheaper than the K3 it replaces and half of Opus 5.5. Highspeed costs the same as Opus. One caveat on sourcing: Moonshot's own pricing page wouldn't render its tables, so these prices come from VentureBeat and the ZenMux listing, which agree.

The benchmarks, in plain English

Terminal-Bench 4.0 gives a model coding tasks to finish in a shell. It is the closest public number to "can it drive a coding agent".

• Moonshot's numbers: K3.1 Pro 65.7%, Opus 5.5 66.4%, GPT-6 Astra 58.2%. Close to Opus, ahead of Astra, and still behind Opus.
• Artificial Analysis, which tests independently, measured Kimi K3 at 13% and K3.1 Pro at 67%. That jump is the post-training at work.
• Artificial Analysis Intelligence Index: 44 → 55, 4th overall, behind Opus 5.5, Sonnet 5.5 and Meta Muse.
• Hallucinations: Moonshot's internal stress test has the claim-level error rate falling from 23.4% (K3) to 8.3%. The prompts are deliberately hard, so read it as a direction, not as how often it will be wrong on your work.

"Beats Opus" is not what the numbers say. It trails Opus slightly on Terminal-Bench, and by 25 points on Terminal-Bench Science.

Calling it

It is live on kimi.ai and the mobile apps, the Kimi Work desktop app, Kimi Code (pick it with /model) and the API, which is OpenAI-compatible:

client = OpenAI(api_key=KEY, base_url="https://api.moonshot.ai/v1")
r = client.chat.completions.create(
    model="kimi-k3.1-pro",
    reasoning_effort="low",  # low | high | max
    messages=[{"role": "user", "content": "Hi"}],
)

What trips people up:

• Thinking is always on. Trying to disable it returns HTTP 400.
• temperature and top_p are fixed. Leave them out.
• Don't assume a default reasoning_effort. The API docs say max, the Kimi Code docs say high. Set it yourself.
• A mistyped model id in Kimi Code falls back silently to K2.8 Preview, without thinking. Check the id.
• The 1M context in Kimi Code needs the Pro plan or above, and Highspeed burns about 2x the quota.
• The API runs on Singapore servers. Enterprise access is off by default.

Open weights, not yet

"It's open source, so I can download it now" is the common mistake. Moonshot has promised the weights by 2026.11.04, and no licence has been named. Kimi K3 shipped under a custom "Kimi K3 License" with commercial conditions, not MIT. Until the weights and the licence are both out, plan as if K3.1 Pro is an API-only model.

On a client project I'd point a cost-sensitive coding agent at kimi-k3.1-pro through the existing OpenAI client and run our own evals before moving anything off Opus, and I wouldn't plan on self-hosting until the licence is published.

The short version

1. Kimi K3.1 Pro (2026.10.23) is Kimi K3 with much more RL post-training: "3T-class", 1M-token context, image and video input.
2. $2 / $10 per million tokens ($0.20 cached), half of Opus 5.5's $4 / $20.
3. Terminal-Bench 4.0: 65.7% on Moonshot's numbers against Opus 5.5's 66.4%; Artificial Analysis measured K3 13% → K3.1 Pro 67%.
4. Thinking is always on, sampling parameters are fixed, and the API is OpenAI-compatible at api.moonshot.ai/v1.
5. Open weights are promised by 2026.11.04, with no licence named. Not out as of 2026.10.

Sources: Kimi blog, Kimi K3.1 Pro · Kimi API docs, K3.1 Pro quickstart · Kimi Code, models · VentureBeat, Kimi K3.1 Pro flirts with open weights leadership · ZenMux, Kimi K3.1 Pro listing · Unite.ai, launch coverage · LLM Gateway, Kimi K3 open weights