Kimi K3.1 Pro is Moonshot AI’s newest model, and it went live on October 23, 2026 at a lower API price than the model it replaces. The open weights are not out yet. Moonshot says they are coming in early November.
Here are the numbers you will probably want first.
- Released October 23, 2026 on Kimi.ai, Kimi Work, Kimi Code and the Kimi API.
- API price: $2.00 per million input tokens and $10.00 per million output tokens.
- 2.8 trillion total parameters, 104 billion active, with a 1 million token context window.
- Open weights are promised by November 4, so you cannot self-host it today.
The takeaway: it is cheaper than K3 and strong at agentic coding, but most of the proof is still Moonshot grading its own homework.
What Moonshot actually shipped
K3.1 Pro is a post-trained upgrade of Kimi K3, which Moonshot released as open weights in July. It is a mixture-of-experts model, so only about 104 billion of its 2.8 trillion parameters work on any given token. That keeps running costs far below what the headline size suggests.
It takes text, images and video through the API. Thinking mode is always on, and you pick a reasoning effort of low, high or max. Max is the default.
There is also a “Highspeed” variant that Moonshot says runs two to three times faster, at double the price.
Kimi K3.1 Pro price: cheaper than K3
Moonshot cut prices while raising performance, which is the opposite of what most labs do. Here is how the three options compare, according to MarkTechPost.
| Model | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| Kimi K3 | $3.00 | $15.00 |
| Kimi K3.1 Pro | $2.00 | $10.00 |
| Kimi K3.1 Pro Highspeed | $4.00 | $20.00 |
Cache hits on K3.1 Pro cost just $0.20 per million input tokens. If your app reuses long prompts, that is where the savings add up. You need a minimum $1 top-up to get API access.
How the benchmarks look, and how much to trust them
Moonshot reports 65.7% on Terminal-Bench 4.0, a test of command-line agent work. It lists Claude Opus 5.5 at 66.4% on the same test, so the two are neck and neck. Independent tracker Artificial Analysis, in a post on X, put K3.1 Pro fourth on its Intelligence Index with a score of 55.
The picture is mixed once you look past coding. Here are some of the figures Moonshot published.
| Test | Kimi K3.1 Pro | Comparison |
|---|---|---|
| Terminal-Bench 4.0 | 65.7% | Opus 5.5: 66.4% |
| SWE-Marathon | 51.7 | Opus 5.5: 46.3 |
| Terminal-Bench Science 0.1 | 38.1 | GPT-6 Astra: 68.1 |
| Hallucination test (internal) | 8.3% of claims | K3: 23.4% |
It wins on long software tasks but trails by 25 points or more on science-heavy terminal work. These scores are self-reported, and some come from Moonshot’s own internal tests. Treat them as a guide, then try the model on your own workload.
Open weights, and the hacking-skills trade-off
Moonshot says it will publish the weights by November 4. The company has not named a license in the coverage I found, so check the model card before building anything commercial on it. The previous release, K3, followed the same pattern: announced first, weights about ten days later.
One detail stands out. For the open release, Moonshot says it removed cyber-related data from training. On its internal hacking benchmark, an unfiltered checkpoint solved 90.5% of tasks, while the shipped model solves 81.0%. That is a company-reported figure, and nobody has verified it yet.
Open-weight models keep closing the gap with paid ones. If that trend interests you, see our look at Mistral Large 4 and its open weights, and at the DeepSeek funding round that shows how much money is chasing the same race.
Who should try it now
If you run coding agents or long automated jobs, the lower price makes K3.1 Pro worth a test this week. The API is OpenAI-compatible, so swapping it in takes only a few lines. One tip from Moonshot: pass back the full assistant message each turn. Dropping the reasoning history can make the output unstable.
If you need to run it on your own hardware, wait. It is not a single-GPU model, and Moonshot recommends clusters of 64 or more accelerators for K3-class models.
Frequently Asked Questions
Is Kimi K3.1 Pro open source?
Not yet. Moonshot says the open weights will arrive by November 4, 2026. Until then it is available only through Kimi apps and the API.
How much does the Kimi K3.1 Pro API cost?
It costs $2.00 per million input tokens and $10.00 per million output tokens. The Highspeed version is $4.00 and $20.00.
Is Kimi K3.1 Pro better than Claude Opus 5.5?
Not across the board. Moonshot’s own figures show a near tie on Terminal-Bench 4.0, a win on SWE-Marathon and big losses on science tasks.
Can I run Kimi K3.1 Pro on my own PC?
No. At 2.8 trillion parameters it needs a multi-accelerator server cluster, even with only 104 billion active.
My take: K3.1 Pro is a smart cost move, and it will be most interesting once the weights are public and outside testers can poke at it. Test it on your own tasks before switching anything important. Until the independent numbers land, keep your current model as the fallback.
Leave a Reply