Published: 26 October 2026
Moonshot AI released Kimi K3.1 Pro on 23 October, and the independent test firm Artificial Analysis ranks it three points behind Claude Opus 5.5, at half the API price. Moonshot says the full weights follow by 4 November. On Artificial Analysis’ numbers, that would make K3.1 Pro the strongest model anyone can download and run on their own servers.
Only two Claude models and Meta’s Muse Wonder Full score higher than K3.1 Pro
Artificial Analysis gives K3.1 Pro 55 points on its Intelligence Index, up from 44 for Kimi K3. Only Claude Opus 5.5, Claude Sonnet 5.5 and Meta’s Muse Wonder Full score higher.
K3.1 Pro is a retrained version of Kimi K3, which Moonshot released in July. It has the same size: 2.8 trillion parameters, of which 104 billion are used for each piece of text it produces. It reads up to one million tokens at once, enough for a large codebase or a stack of contracts. Moonshot reports the biggest gains in agent work: writing code, using tools and fixing bugs over many steps.
On work and coding tests, K3.1 Pro trails Opus 5.5 by a few points, wins one of Moonshot’s long coding tests and costs half as much per token.
| Measure | Kimi K3.1 Pro | Claude Opus 5.5 |
|---|---|---|
| Intelligence Index (Artificial Analysis) | 55 | 58 |
| GDPval-AA office tasks, Elo rating (Artificial Analysis) | 1,846 | 1,866 |
| Terminal-Bench 4.0, coding in a terminal (Moonshot) | 65.7% | 66.4% |
| SWE-Marathon v1.1, long coding tasks (Moonshot) | 51.7 | 46.3 |
| API price per million tokens, input and output | $2 and $10 | $4 and $20 |
Artificial Analysis measured the first two rows itself, Moonshot ran the coding tests, and VentureBeat reported Anthropic’s Opus price.
Treat the praise on social media with some care. VentureBeat reported that an X user shared screenshots of what he called instructions for a paid Kimi campaign. They told creators not to label posts as paid partnerships and to present Kimi as an open-source Opus 5.5 at a fraction of the price. The document has not been independently authenticated, and VentureBeat found no response from Moonshot.
Moonshot removed hacking data to keep the open model behind closed rivals
Moonshot noticed that the model got better at exploiting software flaws as training went on. For the open release, it removed cyber-related data from two training stages and kept the general coding tasks. On Moonshot’s internal exploit test, a version without that filter solved 90.5% of the tasks and the released model 81%.
Moonshot says open weights should not push the frontier of attack capability, and that K3.1 Pro stays behind closed models that have been public for at least three months. Security teams that want a model for penetration testing should look elsewhere first.
Where does your data go when you use Kimi K3.1 Pro?
To servers in Singapore: the privacy policy of Moonshot’s API platform says it stores the data it collects there. The policy has a section for users in the European Economic Area and promises safeguards for transfers abroad, but it names no European data region. API customers can ask for zero data retention, Moonshot says.
K3.1 Pro is available now in these places:
- Kimi.ai, the Kimi apps and the Kimi Work desktop app: free with limited use, and included in paid plans from $15 a month on annual billing.
- Kimi Code, Moonshot’s coding agent: from the same $15 Plus plan.
- The API as kimi-k3.1-pro, or as kimi-k3.1-pro-highspeed at twice the input and output price, which Moonshot says runs two to three times faster.
- Kimi Enterprise: switched off until an administrator turns it on.
Self-hosting K3.1 Pro needs data-centre hardware and a licence that is not out yet
The weights are what make this release interesting for European organisations. A model you run yourself, or through a European hosting partner, keeps prompts and documents out of Singapore. But the weights were still not out on 25 October, and Moonshot has not published the licence.
Simon Willison noted in July that K3’s licence requires a separate deal from companies that sell the model as a service. That applies once their revenue passes about $20 million a year. K3’s files take up 1.56 TB, and K3.1 Pro has the same size, so this is a model for a data centre or a hosting partner, not for a workstation. China’s draft export rules of 16 October also cover the weights of large models, with a registration route for open-source releases. Moonshot has not said whether that affects the 4 November date.
Test K3.1 Pro on coding work now, and decide on self-hosting after 4 November
If your team runs coding agents or document work on Opus 5.5, test K3.1 Pro side by side through the API. Use code and documents that are allowed to leave the EU. Halving the token bill adds up when agents run for hours. Keep personal and confidential data out of those tests until a data processing agreement covers the transfer to Singapore.
Organisations that want sovereign AI should wait for the weights and the licence. Then compare K3.1 Pro with Meta’s Muse Spark Open and Mistral’s models, on their own hardware or with a European host. For the local and sovereign AI projects DataNorth builds, licence terms, hardware cost and who runs the servers usually weigh more than a three-point benchmark gap. Whatever you decide, ask staff not to paste company data into the free Kimi app.
For more information, visit the official announcement of Kimi K3.1 Pro on the Kimi blog.
