Moonshot AI, a Chinese lab, released Kimi K3.1 Pro in late October 2026. On Artificial Analysis's Intelligence Index, which combines many tests into one score, it landed three points behind the best American model. "No relation to this podcast," host Peter Diamandis said of the lab's name on Moonshots with Peter Diamandis, recorded October 27. The panel spent less time on the ranking than on what the narrowing gap means. They asked whether China had copied its way there, whether it now beats the US on value for money, and whether a model anyone can download could ever lead.
What Moonshot released
Diamandis, founder of XPRIZE and Singularity University, read out the specifications. K3.1 Pro is a giant model with 2.8 trillion parameters, the adjustable numbers a model learns during training. Only 104 billion of them are active for any given piece of text, which keeps it cheaper to run. It can take in up to a million tokens at once, the word fragments models read and write. It scored 55 on the index, fourth in the world. Artificial Analysis's launch snapshot of October 23 confirms those figures. It places Kimi behind Anthropic's Claude Opus 5.5 at 58 and two models tied at 56, Claude Sonnet 5.5 and Muse Wonder Full.
Diamandis said the price, $2 per million input tokens and $10 per million output tokens, was "a third of the last Kimi model." Artificial Analysis gives the same prices but describes them as 33% below Kimi K3, a cut of one third rather than a price one third as high. The new model also writes more: about 55,000 output tokens per index task, against 48,000 for K3. As a result, Artificial Analysis estimates the cost per task stayed at about $1.98.
Diamandis also said Moonshot planned to release the weights "next week," meaning anyone could download and run the model. Moonshot's launch post said full weights would follow by November 4, 2026. It describes K3.1 Pro as a further training of Kimi K3 that used more compute for reinforcement learning, a method that improves a model by rewarding good results.
Diamandis framed the release as a milestone for open models. He said the gap between the best open and closed models had shrunk from 12 points to 3, and cited Ramez Naam's estimate that Kimi is now about five weeks behind the US frontier. "That's the shortest open-closed AI gap has ever been," Diamandis said.
The distillation charge
The conversation turned straight to distillation, which means training one model on the answers of a stronger one so that it learns to imitate it. The claim put forward was that K3.1 Pro is plainly a distillation of Opus 5.5, judging by what it can do. Distillation, the argument went, also leaves odd fingerprints. As an example, the model was said to write 3D files for games directly, without working through tools such as Blender or Unreal, supposedly because it had learned how Opus 5.5 represents such objects.
Moonshot's own 3D results are more mixed. Its examples include scripts that run inside Blender or computer-aided design software. On BenchCAD, which tests rebuilding 3D objects as CAD code from images taken from several angles, the company reports Kimi at 92.7% with tools, against 96.2% for Opus 5.5 and 95.9% for OpenAI's GPT-6 Astra. The three scores do not come from one shared test run. Moonshot ran open models itself on 1,000-case random samples of the full 17,900-case dataset. It took the Opus 5.5 score from Anthropic's system card and the GPT-6 Astra score from OpenAI's official report.
The distillation charge has a documented source. In its September 2026 threat report, Anthropic alleges that Moonshot used fraudulent accounts to route some Kimi customer requests to Claude, keep the responses and extract Claude's reasoning. Anthropic attributes more than 23 million exchanges to the campaign between May and July 2026. In one ten-day stretch it counted almost 300,000 relayed requests. These are Anthropic's investigative findings and allegations.
No one on the panel argued that distillation had not happened. Alexander Wissner-Gross, a computer scientist and founder of Reified, said that "based on allegations," distillation was "probably a large portion of the story." Dave Blundin, founder of Link Ventures, agreed ("Sure, they distilled for sure") but said he wanted "to stay away from the narrative that they just copied." Moonshot does not have Anthropic's source code, he pointed out. In his view, distillation "gives you a really good roadmap to build toward, but it doesn't give you the answer key," so Moonshot must have innovated on its own as well.
Ahead on value for money?
Diamandis noted that Kimi sat at the top of Humanity's Last Exam. Wissner-Gross waved that away: "Humanity's last exam is cooked, long cooked." The test, introduced in January 2025 by Long Phan and colleagues, has 2,500 closed-ended questions across more than 100 academic subjects, chosen because they were hard for the models of the time. Its authors say that doing well on such questions is not the same as doing open-ended research. Artificial Analysis has also dropped the original version from version 5 of its index, which its version history places in October and November 2026, alongside a rebalanced mix of newer tests. Scores from different index versions measure different mixtures of tests, so gaps measured months apart do not sit on a fixed scale.
Wissner-Gross argued that the more important result was cost. Measured as performance per cost per task, Kimi "appears, at least to my eye," to have passed Opus 5.5, which he called "extraordinary." In that "maybe spiky slash subjective sense," he said, China may have passed the American public frontier. He welcomed it. Strong Chinese models with open weights, he said, help keep "anti-competitive AI safety cartels in the U.S. at bay," forcing American labs to compete and "making them dance." He called that a net positive for the US and the world.
A history without Chinese pressure
Blundin went further and imagined the world without Chinese competition. In what he called "Dario's fantasy world," a reference to Anthropic CEO Dario Amodei, the company "would have stopped at Fable 5.1." Blundin said Amodei did not even want to release that model and did so only because of competitive pressure. He described Fable 5.1 as the first model capable of improving itself, and said Kimi is now "clearly capable of self-improvement" and available to anyone.
That, he said, changes the plans of "every large corporation with any pride whatsoever, every sovereign nation with any pride whatsoever," because Kimi's cookbook is open. A country or company that gets a foothold with such a model can use it to improve itself, provided it takes the work seriously. Without China, Blundin said, AI would be "completely behind closed doors." Three or four providers would accelerate their models internally without releasing them, and you would have to join a company to get access. "The Chinese have changed the course of history," he said. He added that the trend line looks "shockingly consistent," with no reason to expect the next step to stall.
What 55 means in practice
Blundin pressed the group on a plainer question: what does a score of 50, 60 or 70 actually feel like, since 60 and 70 are "coming very, very soon"? The rule of thumb offered in reply was that 50 is a decent employee you have to keep an eye on, 60 an incredibly competent one, and 70 a superstar. Artificial Analysis itself describes its index as a composite whose makeup changes between versions, not a fixed measure of intelligence.
Could open models take the lead?
Diamandis asked whether open models could ever exceed closed ones. "Of course," said Salim Ismail, founder of Open ExO. Wissner-Gross said yes, in a world where Western frontier labs keep their best capabilities internal for legal or anti-competitive reasons, which he said may or may not be where things are headed.
Blundin disagreed. "Kind of by definition," he said, if open models overtake closed ones, the closed labs will simply use the open models, so open models cannot pass the best internal systems. Once models reach the level of a highly competent or genius employee, he said, whoever has the most compute has a serious advantage. He gave Anthropic's IPO and $100 billion raise as an example of the advantage of controlling a huge amount of compute.
Wissner-Gross answered with an analogy from software. He argued that the Linux kernel, the core of the open-source operating system, is cleaner in its architecture than the Windows kernel, while granting that "folks may throw tomatoes and disagree." That holds, he said, even though Linux source code has been open for decades for any Microsoft engineer to read, and Microsoft has enormous resources. Legal, cultural and architectural reasons mean Microsoft can see Linux but cannot simply use it. Open-weight models could end up in the same position, he said: "possible, not saying it's likely." Frontier labs would see the open models through their window but could not fold them into their own training, the final tuning stage known as post-training. He put it as a twist on Coleridge's line about the becalmed sailor: "Water water all around, but not a drop to drink. Models, models all around, but not a post-training run to distill."