New Chinese AI model cuts costs

Developers building software with artificial intelligence can expect lower bills this month. A new model from a Chinese company now ranks just behind the industry leader. It also costs less to run. This change matters for any business using AI to write code or handle tasks.

Moonshot AI released Kimi K3.1 Pro on October 23, 2026. It is an updated version of the previous Kimi K3 model. The new version uses more compute during the training process. This extra effort has led to significant improvements in performance.

The model has climbed to fourth place on the Artificial Analysis Intelligence Index. It scores 55 points on that leaderboard. Only three models rank higher than it. These are Claude Opus 5.5, Claude Sonnet 5.5, and Muse Wonder Full. The gap between the new model and the top spot is only three points.

The biggest jump is in agentic tasks. These are actions where the AI performs multi-step jobs. On the GDPval-AA benchmark, the model scores 1,846 Elo. That is 312 points higher than the previous version. It also performs well on coding tests. On Terminal-Bench 4.0, it achieves a 65.7 percent success rate. This beats GPT-6 Astra and is close to Claude Opus 5.5.

The model makes fewer mistakes. Internal tests show a drop in hallucinations. These are errors where the AI states false facts. The rate fell from 23.4 percent to 8.3 percent. Moonshot notes this is a stress test. It is not a measure of general error rates across all uses.

The company also improved the model for 3D work. It can work with game engines like Unity and Unreal. It can handle both quick fixes and full project builds. On the BenchCAD benchmark, it scores 92.7 percent with tools. This is close to the top models. Moonshot says it is strongest at modeling mechanical parts and architecture.

The model size stays the same. It has 2.8 trillion total parameters. But the price has dropped. Input costs are now two dollars per million tokens. Output costs are ten dollars. This is a 33 percent reduction from the previous version. The context window remains at one million tokens.

Moonshot plans to release the model weights by November 4, 2026. If this happens, it will be the top open-weights model. This would shrink the gap between open and private models. The difference would drop from 12 points to just three. This makes the technology more accessible for many users.

The company also addressed cybersecurity. It removed cyber-related data from the training set for the open version. This limits its offensive capabilities. The cleaned model solves 81.0 percent of hard security tasks. The uncleaned version solved 90.5 percent. Moonshot wants to balance power and safety. It believes open models should serve both offense and defense.