Kimi K3.1 Pro: Scores, Price and Limits

01What Kimi K3.1 Pro is and how it works

Kimi K3.1 Pro is an open-source AI model from a Chinese company called Moonshot AI. On the hardest test for AI agents it scored 65.7%, while Claude Opus 5.5 scored 66.4%.

That is less than one point between the best AI Anthropic sells and a model that anyone will be able to download from November 4th. It also costs about half as much as Opus to use.

It is a huge AI brain with 2.8 trillion parts, but it only switches on about 104 billion of them for each word it writes. It is giant, but it does not use the whole thing every time. That is how it stays fast and cheaper to run.

It can hold 1 million tokens at once, which is roughly three quarters of a million words. It can also see pictures and videos, not just text.

02Why Kimi K3.1 Pro shrinks the gap with US models

For the last few years, the very best AI came from a few American companies: Anthropic, OpenAI and Google. Chinese models were the cheaper option, good but a step behind. Anthropic CEO Dario Amodei once said Chinese AI was maybe six to 12 months behind.

Kimi K3.1 Pro makes that gap look much smaller. A group called Artificial Analysis tests models on the same set of tests and gives each an overall score.

• Kimi K3.1 Pro scored 55.
• Claude Opus 5.5 scored 58.

Before this, the best open model was 12 points behind the best closed model on that same score. Now it is only three. In one release, the gap went from 12 down to three.

Artificial Analysis puts it fourth overall, behind Opus 5.5 and two other models. On a ranking where real people vote on which AI builds better websites, it is eighth.

03How Moonshot improved Kimi K3.1 Pro so fast

Moonshot released Kimi K3 in July. Three months later came Kimi K3.1 Pro, built on the same base brain. They did not build a new one.

What they did was give it far more practice. It did thousands of practice jobs across hundreds of different areas, with a reward every time it got something right. Julian Goldie calls it the same body with way more time in the gym.

The jump from that practice is huge.

• On the same agent test, K3.1 Pro scored 65.7%, while Moonshot says the old K3 scored about 13%.
• On the website-building ranking, Kimi went from 17th place to 8th.
• Moonshot also dropped the price by about a third compared with K3.

So Moonshot made it better and cheaper in the same release.

04Kimi K3.1 Pro benchmark results against American models

Julian runs through the numbers against the American models. On several tests built around long jobs and hard research, Kimi came first.

• Marathon tests with many steps: Kimi scored 51.7, Claude Opus 5.5 scored 46.7, and GPT-6 Astra scored 42.5.
• Humanity's Last Exam has 2,500 of the hardest expert questions people could write. With tools, Kimi scored 70, Opus 67.7 and GPT-6 Astra 57.2.
• Agent's Last Exam is 122 tough questions an agent finishes on its own. Kimi scored 43.2, Opus 42.6 and GPT-6 Astra 32.8. An agent that just looked up the answer online got a zero, so the score is for real work.
• BrowseComp tests how well an agent hunts down hard-to-find answers on the web. Kimi scored 92.1, GPT-6 Astra 91.5 and Opus 90.8.

On long agent jobs and hard research, a Chinese open model is now at the top of the chart.

05Kimi K3.1 Pro price compared with Claude Opus 5.5

Through Moonshot's API, Kimi K3.1 Pro costs $2 for every million words you send in and $10 for every million words it writes back. Tokens are not exactly words, but that is close enough.

Claude Opus 5.5 costs $4 in and $20 out. So Kimi is about half the price of Opus. GPT-6 Astra is reported at about $10.

Moonshot showed one side-by-side to make the point. Both models got the same job: build a 3D screen of a rocket launching.

• Kimi finished in about six minutes for $1.95.
• Opus took about 24 minutes and cost about $10.35.

Opus's rocket looked a little cleaner in places, but Kimi was almost as good for a lot less.

There is a catch. In some tests Kimi is not the cheapest per job, because it can use a lot of words while it thinks. The price on paper and the price of a finished job are not always the same.

06What Kimi K3.1 Pro can do: long jobs, fewer errors and 3D

The best example of how it thinks is a job Moonshot gave it. It had to train a body to walk. At the start, the body collapsed within seconds, so Kimi kept working over 167 missions and 420 reviews.

By the end the body walked on its own for 11.2 seconds. The arms kept moving wrong, so Kimi stopped tweaking the rewards and tested the simulator itself. It found 36 errors, fixed the bad maths, and trained the walk again.

Julian compares that to a senior engineer who checks whether they are fixing the wrong thing.

It also makes up fewer facts. On a test built to trick AI, Opus 5.5 hallucinated 15% of the time and Kimi 8.4%, down from 23.4% for K3. GPT-6 Astra did best at 5.5%.

This matters because agents work in chains, and one bad fact early can ruin every step after it.

Kimi also works in real tools like Unreal Engine, Unity, Blender and CAD by writing its own small scripts. Moonshot showed a palace in Unreal Engine, a full action game with a two-part boss fight, a jet engine built part by part and a lamp ready to 3D print.

07Where Kimi K3.1 Pro falls behind

Moonshot is clear about weak spots, and they are big ones.

• Terminal Bench: a test of science-heavy computer jobs. Kimi scored 38.1 and GPT-6 Astra scored 68.1.
• ProgramBench 2: building four programs that pass every check. Kimi scored 31.5 and Opus scored 65.
• Using a computer like a person: Kimi is a couple of points behind Opus.

Also remember that the numbers are self-reported by Moonshot. Some tests are public and some are their own. Outside groups still rate it strongly.

You might think you will just run it yourself, but you probably will not. Even shrunk down, the files come to about 1.5 terabytes. You would need a very large server full of high-end chips, so this is for big companies, not a laptop.

08Who Kimi K3.1 Pro is for and why it matters

For most people, open source means something else. Lots of hosting companies can offer the model and compete on price. That can bring costs down for you.

It also means you can swap a cheap model like Kimi into the jobs where it makes sense, and keep a top model for the jobs where it does not. Kimi looks strongest on long agent jobs and hard research. It looks weaker on terminal work and building programs.

If you use AI agents in your business, this is the main takeaway. The best brain is no longer only from three American companies, and the price of a very strong model has dropped.

Julian says it is worth learning how to plug Kimi into your own setup while it is new. Inside his community, he adds Kimi to the Agent OS next to Claude, Hermes and OpenClaw.