Chinese company Moonshot AI has unveiled Kimi K3.1 Pro, a powerful new artificial intelligence model that it presents as a genuine rival to the most expensive closed systems in the world, yet one that developers can inspect, run and build upon far more freely. The release is the first in the company’s new K3.1 line and, in Moonshot’s telling, its most capable open-weight model to date, built for serious coding, long research tasks and work that stretches across many steps without losing track of the goal.
At the heart of the announcement is a striking claim about performance. In the company’s own testing, K3.1 Pro scored 65.7 percent on Terminal-Bench 4.0, a demanding test of real-world coding ability, putting it ahead of OpenAI’s GPT-6 Astra at 58.2 percent and only a fraction behind Anthropic’s Claude Opus 5.5 at 66.4 percent. The model can also take in an enormous amount of material at once, with a context window of one million tokens, roughly the equivalent of hundreds of thousands of words, which lets it work through whole codebases, long reports and sprawling research files in a single conversation.
Photo: Wikimedia Commons
Moonshot is careful to describe K3.1 Pro as an evolution rather than a clean break. It is a heavily retrained version of Kimi K3, the 2.8-trillion-parameter model launched in July, put through large-scale reinforcement learning across hundreds of fields at the same time. The company says it rebuilt parts of its training and testing setup to make that possible, from how the model reasons over long passages to how its practice tasks are scheduled, and that the effort produced steady gains in every area it measured. Alongside the main model comes a faster sibling, K3.1 Pro Highspeed, which runs two to three times as quickly, and both are already available through Kimi.ai, Kimi Work, Kimi Code and the Kimi programming interface, with the full model weights promised by 4 November.
The benchmark sheet Moonshot published is, as always with a company’s own numbers, best read with a pinch of salt, but it is detailed enough to be interesting. K3.1 Pro leads its rivals on several tests, including a long coding marathon where it scored 51.7 against 46.3 for Claude Opus 5.5, and a difficult general-knowledge test taken with tools, where it reached 70.0 against 67.7. It also comes out on top in tests of office work, web research and automation, and it edges past GPT-6 Astra on a ranking of workplace tasks while remaining just behind Claude Opus 5.5. Against other openly available models the gap is far wider: DeepSeek’s contender scored 42.3 and another rival 41.8 on the same coding benchmark, more than twenty points behind Kimi.
Yet the picture is far from a clean sweep, and Moonshot’s own figures show where the model still lags. On a science version of the coding benchmark it managed only 38.1, well behind both GPT-6 Astra at 68.1 and Claude Opus 5.5 at 63.3, and it trails the leaders on several other respected tests of programming, mathematics and science. In cybersecurity tests it also sits clearly behind its closed rivals, something the company says is deliberate rather than accidental. Moonshot explains that it stripped cyber-attack material out of the model’s training for the open release, accepting a lower score on its own hacking test, where the finished model solved 81.0 percent of tasks against 90.5 percent for an unrestricted version, because it does not want the freely shared weights to push forward the ability to find and exploit software flaws.
Beyond the headline scores, the release shows how far these models have moved into practical, hands-on work. Moonshot says K3.1 Pro can take on anything from a quick repair inside an existing codebase to building an entire project from scratch, working with popular game engines and combining web search with three-dimensional generation to create complete experiences. In one internal experiment, it built a walking simulation on top of an open anatomical model with 296 skeletal parts and 402 muscles, gradually coaxing a figure that collapsed within seconds into one that walked continuously for 11.2 seconds, and along the way it tracked down and corrected 36 errors in the shoulder-muscle calculations before trying again. The model can also turn a written brief, a sketch or a reference picture into a usable three-dimensional object, scoring 92.7 percent on a test of rebuilding objects from images, close to both leading closed models, and a new 3D Make feature lets users preview, edit and download models for printing.
Finance is another area where Moonshot claims a leap forward, with the model able to carry a complete financial workflow from start to finish and scoring 58.6 on a frontier finance test, ahead of both Claude Opus 5.5 and GPT-6 Astra. The company also reports a sharp fall in invented facts, with the rate of false claims on its deliberately tricky internal test dropping from 23.4 percent for Kimi K3 to 8.3 percent for K3.1 Pro, better than Claude Opus 5.5 but still behind GPT-6 Astra. For developers, the most immediate difference may simply be the bill: access through the programming interface costs two dollars per million input tokens and ten dollars per million output tokens, a third less than Kimi K3, with cached input at just twenty cents, and no extra charge for using the full million-token window.
Taken together, the launch underlines how quickly the open side of the AI race is closing on the closed giants, and how much price pressure that creates for everyone else. If Moonshot’s numbers hold up in independent testing, a model that anyone can download and adapt will soon do work that until recently required the most expensive systems on the market, at a fraction of the running cost. The honest caveats remain, however: these are the maker’s own results, measured at maximum effort, with real weaknesses in science and security testing, and the open weights themselves will not arrive until early November. What is already clear is that the contest is no longer only about who builds the smartest model, but about who can put that power into the most hands at the lowest price, and on that measure Kimi K3.1 Pro has just raised the bar considerably.
