Moonshot's Kimi K3.1 Pro lands within a point of Claude Opus 5.5 on a top coding test, and the weights are coming

Moonshot AI released Kimi K3.1 Pro on Friday, Oct. 23, 2026, the first model in its K3.1 series, with company-reported benchmark scores that put it close to the top proprietary models on coding and agent tests.

The gap between the best open-weight model and the best closed one keeps shrinking, and Moonshot is now claiming to be within a point of Claude Opus 5.5 on a hard terminal-coding test and ahead of GPT-6 Astra. If the weights really ship by Nov. 4, anyone can run something near the frontier on their own hardware, which is great for builders and awkward for anyone charging premium prices for proprietary APIs. The skeptic's note: these are Moonshot's own benchmarks, several are internal, and "close to" is doing some heavy lifting. Also worth noticing: a Chinese lab voluntarily dialing down its model's hacking skills before release. That's either responsible or very good PR. Maybe both.

Moonshot AI introduced Kimi K3.1 Pro, the first model in its new K3.1 series. The company calls it its most capable 3T-class open-weight model. 3T-class means a model in the neighborhood of 3 trillion parameters, the internal settings a model learns in training. Open-weight means the company plans to publish those files so someone else can run the model, rather than keeping it only on Moonshot's own site. It has a 1-million-token context window. A token is a chunk of text, often a word or part of a word, so a million tokens is a very long document the model can hold in one go. It is a post-trained version of Kimi K3. Post-training is extra practice after the first big training run. Moonshot says this version used substantially more of that practice, and that the method was reinforcement learning, which scores the model's attempts and trains toward the better ones. Those lines are Moonshot's.

Moonshot also released Kimi K3.1 Pro Highspeed, which it says runs two to three times faster. Both models are available now on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API. An API is the door a program uses to call the model. Moonshot says the full model weights, the files that store what the model learned, will be released by Nov. 4. Those lines are Moonshot's.

The scores that follow are Moonshot's own. On Terminal-Bench 4.0, a test of coding work done in a terminal, the text window where someone types commands, Kimi K3.1 Pro scored 65.7 percent. Moonshot lists GPT-6 Astra at 58.2 percent and Claude Opus 5.5 at 66.4 percent on the same test. 65.7 is 0.7 points under the 66.4 the company prints for Opus, and 7.5 points above the 58.2 it prints for GPT-6 Astra. Those comparisons are Moonshot's.

On BenchCAD, a test of rebuilding a 3D shape by writing CAD code, the model scored 92.7 percent with tools. CAD is computer-aided design, the software people use to draw parts. Moonshot lists Opus 5.5 at 96.2 percent and GPT-6 Astra at 95.9 percent. Those figures are Moonshot's.

On OfficeQA Pro v2, a test of reading financial documents, Moonshot reports a score of 64.4 and calls that state of the art, its phrase for the best result it knows of. On Kimi Code Bench 3.0, a long-horizon coding test the company wrote and ran itself, the score is 61.2 percent, against 55.95 percent for K3. Long-horizon means the job takes many steps. Internal means other labs did not score this test. Those figures are Moonshot's.

On Moonshot's internal hallucination test, called Kimi Hallu Bench Hard, the claim-level hallucination rate fell from 23.4 percent for K3 to 8.3 percent. A hallucination, here, is a confident statement that is not true. Claim-level means the score counts individual claims, not whole answers. That test is Moonshot's.

Moonshot says it analyzed the model's vulnerability-research capability, the skill of looking for security holes, and applied targeted training controls. It says it identified and removed cyber-related data from its mid-training corpus, the pile of text used in the middle of training, and from its reinforcement-learning mix, to reduce the risk of misuse. Those lines are Moonshot's.

Moonshot says its new 3D Make feature is now live on kimi.ai and the Kimi app. The picture is Moonshot's own screenshot of that 3D Make workspace, from the K3.1 Pro launch post.

In plain terms, Moonshot said on Friday that Kimi K3.1 Pro is out, that its own terminal-coding score sits just under the number it lists for Claude Opus 5.5 and above the number it lists for GPT-6 Astra, and that the weights are promised by Nov. 4. The benchmark numbers are the company's. Several of the tests are internal.