Kimi K3.1 Pro Hits 4th on the Intelligence Index at $2/$10

Kimi K3.1 Pro lands fourth on the Intelligence Index at $2/$10 — and it makes the open-weight gap look small

On October 23, 2026, Moonshot AI released Kimi K3.1 Pro, the first model in a new K3.1 line: a 2.8-trillion-parameter mixture-of-experts model with 104 billion active parameters, a 1,048,576-token context window, and text-plus-image input. A faster sibling, Kimi K3.1 Pro Highspeed, shipped the same day at two to three times the output speed. Both are live on Kimi.ai, Kimi Work, Kimi Code and the Kimi API. Full model weights are promised by November 4, 2026.

The headline is not that it got better. It is that it got substantially better at the same parameter count, for a third less money.

The third-party number that matters

Vendor benchmarks are marketing until someone independent runs them. Artificial Analysis put Kimi K3.1 Pro at 55 on its Intelligence Index, fourth of 215 models evaluated. Above it sit only Claude Opus 5.5 (58), Claude Sonnet 5.5 (56) and Muse Wonder Full (56). It lands ahead of OpenAI's GPT-6 Astra (53) and Google's Gemini 4 Argon (53). Its predecessor, Kimi K3, scored 44 — an 11-point jump with no change to the base model.

Artificial Analysis also stated the consequence plainly: once Moonshot publishes the weights, K3.1 Pro becomes the highest-scoring open-weights model available, and the gap between the best open and best proprietary model on the index narrows from 12 points to 3.

That is the real story. For eighteen months the standard objection to open models has been that they are cheap but fall apart on long-horizon work. A three-point gap does not eliminate the objection, but it moves the argument from capability to operations: hosting cost, latency, compliance.

Where the gains actually are

The index score is an average, and the average hides a very uneven profile. The improvements are concentrated almost entirely in agentic work.

GDPval-AA (real-world work tasks): 1,846 Elo, up 312 points from K3. Third overall, behind Claude Opus 5.5 (1,866) and Muse Wonder Full (1,857).
Terminal-Bench 4.0 (agentic coding in a terminal): 67.2%, the highest score of any model, proprietary or open. Kimi K3 scored 13%. This is the single largest movement in the release.
AutomationBench-AA (agentic SaaS workflows): 58% to 73%.
Elsewhere gains are moderate: Humanity's Last Exam 47% to 52%, SciCode 59% to 61%, CritPt 23% to 25%, and the AA-Omniscience index 20 to 30 with hallucination falling from 53% to 44%.

And where they are not:

AA-LCR v1.1 (long-context reasoning): 86.3%, below Kimi K3's 88.7%.

Pricing: $2 / $10 / $0.20, and the catch

Kimi API pricing, per million tokens, excluding tax:

kimi-k3.1-pro — $2.00 input, $10.00 output, $0.20 cache hit
kimi-k3.1-pro-highspeed — $4.00 input, $20.00 output, $0.20 cache hit
Cache writes — $2.00 (5-minute TTL) or $4.00 (1-hour TTL); $4.00 / $8.00 in Highspeed

Kimi K3 was $3.00 / $15.00 / $0.30. All three rates drop by one third. There are no context-length tiers: the rate is flat across the full million-token window, which is unusual and genuinely useful for long-document work.

Now the catch. Cost per task is not down. Artificial Analysis measured K3.1 Pro at $1.98 per Intelligence Index task, roughly equal to K3, because it emits 55K output tokens per task versus K3's 48K — 14% more. The 33% unit-price cut is largely cancelled by the token increase. If you budgeted a 33% saving, you will not see it.

Part of that is by design. K3.1 Pro always reasons; reasoning_effort accepts low, high and max, and the default is max. Running ordinary extraction or classification at max burns output tokens for nothing. Dropping to low on simple work is the single cheapest optimization available.

Consumer plans

Kimi's membership tiers use annual billing shown as effective monthly rates: Free (limited K3.1 Pro, no Kimi Code); Plus $15 (K3.1 Pro on web, Work and Code, no Highspeed); Pro $31 (Highspeed, Kimi Claw, Goal mode); Max $79 (million-token chat, 5× agent credits); Ultra $159 (10× credits, 4 concurrent tasks). Free access to the flagship is new relative to K3, but Moonshot only says "limited usage" — the token allowance is unpublished, so budget for Plus if the work matters.

3D and finance are the underrated half of the release

Moonshot spent much of its blog on two domains outside coding, and both are more relevant to non-developers than Terminal-Bench.

3D. On BenchCAD — reconstructing geometry from multi-view renders by generating CAD code — it scores 92.7% with tools, versus Opus 5.5 at 96.2%; without tools, 72.8% against 73.0%, effectively a tie. It writes and runs Blender and CAD scripts without MCP connectors, so parametric modeling is its natural fit. Moonshot also launched 3D Make inside kimi.ai. Finance. It scores 64.4 on OfficeQA Pro v2 — retrieval, cross-document verification and numerical reasoning over real financial documents — and Moonshot claims state of the art ahead of Opus 5.5 and GPT-6 Astra. On its internal Kimi Hallu Bench Hard, claim-level hallucination fell from 23.4% (K3) to 8.3%. All vendor-reported, with no independent reproduction yet.

The cybersecurity decision is the most interesting part

Moonshot did something unusual: it removed capability before release. It traced rising vulnerability-research performance to cyber content in the general mid-training corpus and to cyber tasks in the RL mix, then stripped it from both stages for the open-weight release.

On its internal xdev-bench-hard — tasks reproducing recently disclosed Linux kernel, Node.js and JavaScriptCore vulnerabilities, requiring a working exploit — K3.1 Pro scores 81.0% pass@3. The same checkpoint without data removal scores 90.5%. Moonshot published that counterfactual itself. The stated policy: do not extend the absolute frontier of offensive cyber capability in open weights, and raise it gradually. The trade was roughly ten points of cyber capability for a defensible open-weights position.

The license is not MIT

Weights arrive by November 4 under the Kimi K3 License, not Apache 2.0 or MIT. It reads permissive — use, copy, modify, sublicense, sell, fine-tune — with two gates:

Model-as-a-Service revenue gate. Running a MaaS business (third-party inference or fine-tuning with meaningful control over inputs, parameters or training data) at above $20M aggregate revenue over any consecutive 12 months requires a separate agreement with Moonshot.

UI attribution. Products above 100M monthly active users or $20M monthly revenue must display "Kimi K3" prominently.

Internal use is exempt from both, and so is access through Moonshot's own products or certified partners. A company self-hosting K3.1 Pro for internal search and drafting has no license problem; the gated case is reselling Kimi inference as your own API. Calling this "open source" is wrong — it is open-weight, tagged license: other on Hugging Face, and that distinction decides whether procurement needs legal review.

What to do this week

Three paths:

Test it today. Call kimi-k3.1-pro on the Kimi API with twenty of your own tasks and measure accuracy and token burn. At $2/$10 the experiment is cheap, and the flat million-token window makes long-input tests unusually affordable.
Wait for November 4 to self-host. K3's weights were roughly 1.4TB resident and K3.1 Pro should be similar; vLLM-class stacks that already support K3 are the starting point. This is a data-center decision, not a laptop one.
Use membership for evaluation only. Plus at $15 covers web and Kimi Code; Pro at $31 adds Highspeed for interactive work. Treat free-tier "limited usage" as a demo.

The one thing not to do is build a schedule around November 4. Kimi K3's weights shipped eleven days after the promised date in July. What is certain today is that the API works and the price is published. Everything else is a date on a blog post.

The honest summary: K3.1 Pro does not prove that open-weight models have won. It proves that the capability gap is now small enough that the decision turns on hosting economics, data residency, and whether you can tolerate a model that thinks at max effort unless you tell it not to.