Kimi K3.1 Pro, Moonshot AI's new model, ranks fourth on the Artificial Analysis Intelligence Index, just behind Claude Opus 5.5 (a score of 55 against 58), at roughly half the list price. It is billed as open weight, but the weights are only promised for November 4, so today you can reach it through Moonshot's API and nothing else. Cheaper per token also does not guarantee cheaper per finished task.
What "open weight" means, and why it isn't here yet
Open weight means the model's files are released for anyone to download. Several posts on X already call K3.1 Pro open weight, but Moonshot launched it on October 23 and says the full weights arrive by November 4, as summarized from Reuters by @pbsIdentity. Polymarket had a "Moonshot publishes Kimi K3.1 Pro weights" market at about 90% when we checked. The license is the open question: none has been announced, and the previous model, K3, shipped under a custom license rather than an OSI-approved one. If you run a business, read the license before building on the weights.
Why you can't run it on a laptop
K3.1 Pro is a post-trained K3 with about 2.8 trillion total parameters and a 1M-token context window. Creator write-ups put local running somewhere between roughly 340GB of memory with heavy compression and well over a terabyte without it. Alex Finn's recent video shows what serious local hardware looks like (dual DGX Spark machines for a smaller model). For almost everyone, K3.1 Pro means an API.
Cheaper per token, not always cheaper per task
List price is $2 per million input tokens and $10 per million output tokens, about half of Claude Opus 5.5's rate card, per DataCamp and other pricing guides. A Highspeed tier runs $4 and $20 and is 2 to 3 times faster. One analysis found about 14% more output tokens than K3, and a coding test at low reasoning effort by Atomic Agent cost about half of Opus 5.5 but took roughly twice as long. Nick Saraev's head-to-head and Pat Simmons' full review both found Opus 5.5 ahead on polish, and one tester reported having to step in repeatedly on a 3D billiards build. Every fix-up prompt burns tokens, which can eat the headline saving. Treat that as one tester's experience, not a verdict.
The numbers to treat with care
The hallucination rate reportedly fell from 53% to 44%, but we could only find that figure in secondary coverage, so treat it as reported. Terminal-Bench and similar scores are vendor-published. VentureBeat also reports allegations of paid influencer promotion around the launch, which is unverified. Moonshot says it removed cyber-related training data, and its own internal benchmark showed 81.0% for the shipped model against 90.5% for an unfiltered checkpoint (Technology.org). That is Moonshot's figure and it has not been independently checked.
Try it today: a two-step test
Open an API router such as OpenRouter, pick Kimi K3.1 Pro, and give it one of your recurring daily tasks, such as a spreadsheet formula or routine data entry. Compare the result to what you use now before you cancel anything.
Before you paste any company data, check the provider's data-retention settings and confirm prompts are not stored or used for training.
FAQ
Is Kimi K3.1 Pro open source?
Not yet. It is described as open weight, with weights promised by November 4 and no license announced. Open weight does not automatically mean open source.
How much does Kimi K3.1 Pro cost?
List price is $2 per million input tokens and $10 per million output tokens, about half of Claude Opus 5.5. Cost per finished task can differ because the model may use more tokens.
Can I run Kimi K3.1 Pro on my own computer?
Almost certainly not. Reported memory needs run from roughly 340GB to over 1TB, which means server-class hardware.
Is Kimi K3.1 Pro better than Claude Opus 5.5?
On the Artificial Analysis index it scores 55 against Opus 5.5's 58. Hands-on testers report Opus 5.5 is still ahead on polish and complex spatial tasks.
Companion episode:
Want more plain-English AI breakdowns for solo builders and small teams? Get them free at joebuildsai.com.
