Moonshot AI introduced Kimi K3.1 Pro on 23 October, the first model in the K3.1 line. It is a post-trained K3 with a lot more post-training compute, still 2.8 trillion parameters, with a one-million-token context window. Alongside it comes K3.1 Pro Highspeed, two to three times faster. Both are live on Kimi.ai, Kimi Work, Kimi Code and the Kimi API, with full weights promised by 4 November.
Fourth place, and who is ahead
On the Artificial Analysis Intelligence Index, K3.1 Pro scores 55 points, good for fourth. Only Claude Opus 5.5 at 58, Claude Sonnet 5.5 at 56 and Muse Wonder Full at 56 rank higher. That is an eleven-point jump over K3, unusually large for post-training alone.
The rank matters less than what happens on 4 November. Once the weights land, K3.1 Pro becomes the highest-scoring open-weights model around, and the gap between open and proprietary shrinks from twelve points to three.
The agentic numbers are the story
The biggest gains show up in agentic work. On GDPval-AA, Artificial Analysis puts K3.1 Pro at 1,846 Elo, 312 points above K3 and just behind Opus 5.5 at 1,866. On Terminal-Bench 4.0, Moonshot reports 65.7 percent against 58.2 for GPT-6 Astra and 66.4 for Opus 5.5. K3 managed 13 percent on that same test.
Moonshot’s internal Kimi Code Bench 3.0 Long-horizon goes from 55.95 to 61.2 percent. Claim-level hallucinations on the in-house Hallu Bench Hard fall from 23.4 to 8.3 percent, which Moonshot is careful to call a stress test rather than a general error rate. GPT-6 Astra sits at 5.5 percent there, so the lead holds.
Prices come down by a third: two dollars per million input tokens, ten for output, twenty cents for cached input. Highspeed costs double that.
A note about cyber data
In passing, Moonshot says it stripped cyber-related data out of training phases to cut misuse risk. That is a notable sentence from a lab that then hands out the weights. Once weights are public you cannot bolt safeguards on afterwards, so leaving things out beforehand is the only lever left.
Three points are smaller than they sound
The distance to Opus 5.5 looks tiny on paper, and on Terminal-Bench it is 0.7 percentage points. Still, look at where the numbers come from: part from Artificial Analysis, part from Moonshot’s own measurements, and vendor numbers are rarely the unflattering ones. What is harder to argue with is the direction. In the summer the open gap was a model cycle wide. Now it is a few weeks until a weights download.