Tech • AI • Robotics • Game
It's over... China's New AI Just Caught Up With America's Best
8/10TL;DR
Moonshot AI’s Kimi K3.1 Pro has pushed a Chinese open-weight model to near parity with the top US systems on major agent and coding benchmarks, narrowing a lead that had defined the frontier for most of the year.
KEY POINTS
Near-parity on top benchmarks
On Terminal Bench 4.0, a demanding test of autonomous computer work, Kimi K3.1 Pro scored 65.7%, just behind Claude Opus 5.5 at 66.4% and ahead of GPT-6 Astra at 58.2%. Independent evaluator Artificial Analysis ranked it the fourth-smartest model overall with a score of 55, an 11-point gain over the previous Kimi release.
Open weights sharpen the competitive shift
The model’s full weights are scheduled for release before November 4, allowing anyone with sufficient hardware to run it locally. That would cut the gap between the best freely available model and the best closed model from roughly 12 points to 3, a major change in access to frontier-grade AI.
A stealth debut fueled online scrutiny
Before launch, the system appeared anonymously on Arena under the name Kibitz, where users compared its coding and visual generation against rival models. Internet sleuths linked it to Moonshot through tokenizer behavior, leaked API references, developer tool updates, and code labels pointing to 3D generation, automation, and multi-agent features.
Massive scale and reinforcement learning
Kimi K3.1 Pro is a 3 trillion-parameter model with a 1 million-token context window, built on K3, which launched in July at 2.8 trillion parameters. Moonshot said the key advance was a large expansion of reinforcement learning and a rebuilt training stack designed to turn extra compute into stronger performance across many fields.
Pricing undercuts premium rivals
Moonshot positioned the model below top US competitors on price. K3.1 Pro costs 13 yuan per million input tokens and 68 yuan per million output tokens, with cheaper reuse of repeated input, placing it around half the price of Claude Opus 5.5 and near Claude Sonnet pricing while competing much closer to Opus-class capability.
Strength in coding, agents, and 3D tools
The model is being marketed for bug fixing, full project generation, and autonomous work inside Unity, Unreal Engine, Godot, and Blender. On BenchCAD, it scored 92.7%, close to Opus 5.5 at 96.2% and GPT-6 Astra at 95.9%, and it can generate editable 3D assets directly for tools such as Blender, SolidWorks, and Rhino.
Detailed demos highlighted engineering accuracy
In one test, a hand-drawn aircraft sketch became a full 3D model in about 40 minutes, producing 608 lines of Blender script and dimensions accurate to within 1 millimeter. In another, the model generated a 2,100-part jet engine with cutaways and rotor animations in FreeCAD, underscoring its appeal for design and web-based 3D workflows.
Progress remains uneven outside headline tasks
Some benchmark gains were modest. On Humanity’s Last Exam, performance rose from 47% to 52%; on a physics research test from 23% to 25%; and on a science coding test from 59% to 61%. Early users also reported that larger rivals still show stronger creativity, insight, and writing quality on difficult real-world tasks.
Finance and hallucination rates improved
On Office QA Pro V2, a finance-focused benchmark based on real documents, Kimi scored 64.4, which Moonshot described as the best published result. The company also said hallucination rates on an internal stress test fell from 23.4% to 8.3%, though it noted the test includes many adversarial prompts.
Hacking capability was deliberately reduced
Moonshot disclosed that a stronger internal variant became highly capable at offensive cybersecurity after hacking data entered both pretraining and reinforcement learning. On the company’s exploit benchmark, that version succeeded 90.5% of the time within three tries, versus 81% for the released model after those skills were stripped back to keep the open version behind the cutting edge of attack capability.
Usage caps triggered user backlash
The main complaint after launch was access rather than raw quality. Some paid users reported exhausting a five-hour quota in under an hour, with others saying a single task could consume most of their allowance, raising doubts about how practical the model is despite its benchmark gains.
CONCLUSION
Kimi K3.1 Pro signals that China has re-entered the top tier of frontier AI, not only on cost but on absolute capability. The next test is whether open access, safety tradeoffs, and restrictive usage limits will determine how much that breakthrough matters in practice.
Ask a question