TL;DR. On October 23, Beijing-based Moonshot AI introduced Kimi K3.1 Pro. In the independent index from Artificial Analysis the model scores 55 points, against 58 for the leader, Claude Opus 5.5. Moonshot plans to release the weights by November 4. From then on any company may run the model itself or have it run. Nine days ago the best open model was still 12 points behind the top. The gap between open and closed models is shrinking to three points. But the overall score hides where the model is weak, and cheap per word does not mean cheap per task.
What Moonshot introduced
According to Moonshot's announcement, Kimi K3.1 Pro is a post-trained Kimi K3. According to its model card, the base model has 2.8 trillion parameters, 104 billion of which are active per word, and reads up to one million tokens. Moonshot mainly sharpened long programming work and office tasks with spreadsheets and financial figures.
In the Artificial Analysis index, which combines many individual tests into one number, it jumps from 44 to 55 points. Only Claude Opus 5.5 with 58 and Claude Sonnet 5.5 and Meta's Muse Wonder Full with 56 each rank higher.
Artificial Analysis index, points
closedopenKimi K3.1 Protop, 58 points
Today the model is only available through Moonshot's own service. For the predecessor K3, the weights arrived in July, eleven days after the launch. If they arrive this time too, K3.1 Pro will be the strongest model you can download. When we wrote about Muse Spark Open on October 16, the best open model stood at 46 points. That still holds today.
Half the price per word, not per task
Moonshot charges $2 per million input tokens and $10 per million output tokens, a third less than for K3. According to Artificial Analysis, Opus 5.5 costs twice as much, $4 and $20. But Kimi writes a lot: for the index it produced 150 million output tokens, while the median across all models is 87 million. That is why Artificial Analysis calculates the cost per task.
Cost per task in the index, in dollars
closed modelsKimi K3.1 Pro
Per task, K3.1 Pro costs $2.11. Opus 5.5 costs $5.98 at the highest setting, but only $1.82 at "high", with 54 instead of 55 points. So for Opus the thinking depth you choose matters more than switching to Kimi. If you compare models, do not lay price lists side by side. Run the same tasks instead.
Our analysis: where the overall score misleads
An index is an average. For a business, what counts is how the model performs on its own tasks. Vals AI tests models independently by domain. On October 25 we pulled out the ranks of K3.1 Pro in five of these tests.
Rank of Kimi K3.1 Pro in Vals AI domain tests
top third of the rankingKimi K3.1 Pro in the top thirdKimi K3.1 Pro further back
In terminal programming K3.1 Pro ranks third of 55 models, in the overall index eighth of 27. On tax cases and finance tasks it lands mid-table, on legal questions at rank 22 of 28. If you want to pre-check contracts or sort tax questions, the headline number tells you little.
The catch
- Not open yet. Until the weights are out, every request goes through Moonshot. The broker OpenRouter offers no processing within the EU for this model. Moonshot has not yet named the license for K3.1 Pro.
- Not for the server room. The weights of K3 alone weigh 1.56 terabytes. The vLLM serving recipe names eight Nvidia GB300 with 288 gigabytes each as the minimum, and several such machines for real production traffic.
- Its own test bench. Many figures in Moonshot's announcement come from internal tests. On the programming test Moonshot measured 65.7% and took the competitors' scores from blogs and leaderboards. Vals measures 63.6% independently.
- Deliberately held back. For the open version Moonshot removed training data on cyberattacks. In its own attack test the score dropped from 90.5 to 81.0 as a result. That is intended. If you plan to use the model for security testing of your own systems, you deliberately get less.
What this means for mid-sized companies
For data that should not leave your company, there will soon be an option close to the top, run by a data center of your choice. The predecessor shows how wide that choice can get: through the broker OpenRouter alone, 21 providers offer K3 today, including Amazon Bedrock, Together and DeepInfra.
More important than who is ahead right now is the pace. The best open score rose from 46 to 55 points in nine days. If you tie your application firmly to one model, you pay with every switch. If you treat the model as a replaceable part, you take every improvement with you, whether it comes from California or from Beijing.
What you should do now
- Do not choose by the overall score. Look for the domain tests that come closest to your tasks. Better still: test with 30 of your own non-critical tasks.
- Calculate per task. Run the same tasks through two or three models and compare the bill, not the price list. Try the lower thinking setting as well.
- Wait for the weights. Do not send company data to a service whose processing location and terms you have not checked. After November 4, ask your cloud provider whether and where it runs the model.
- Keep the model replaceable. With a layer between application and model and a fixed set of your own test tasks, a switch becomes a setting instead of a project.
- Read the license as soon as it is out. For the predecessor, internal use is free of conditions. Whether that applies to K3.1 Pro is not yet clear.
Kimi K3.1 Pro shows how fast the lead of closed models is melting. That does not make the choice easier, only larger. If you would like to compare properly which model fits which of your tasks, feel free to get in touch.
