Chinese AI model cuts costs by third

Kimi K3.1 Pro, a new reasoning model from Chinese company Moonshot AI, has reached 55 points on the Artificial Analysis Intelligence Index. This score places the system fourth in the global ranking, just three points behind the leading proprietary models. The result represents an 11 point improvement over its predecessor, Kimi K3.

Only three systems currently outperform the new release. Anthropic’s Claude Opus 5.5 leads with 58 points, followed by Claude Sonnet 5.5 and Muse Wonder Full, both at 56 points. Moonshot plans to publish the model weights by November 4, 2026. If this schedule holds, K3.1 Pro would become the highest-scoring open-weights model available. That move would narrow the performance gap between open and proprietary systems from 12 points to three.

The most substantial gains appear in agentic tasks. On the GDPval-AA benchmark, the model achieved an Elo rating of 1,846. That figure is 312 points higher than the previous version and trails only Claude Opus 5.5 and Muse Wonder Full. In terminal coding tests, K3.1 Pro recorded a 65.7 percent score on Terminal-Bench 4.0. This result exceeds GPT-6 Astra’s 58.2 percent and narrowly misses Claude Opus 5.5’s 66.4 percent. By comparison, the older K3 model managed only 13 percent on that specific test.

Moonshot also reports a steep decline in hallucinations. On its internal Kimi Hallu Bench Hard, claim-level errors dropped from 23.4 percent in K3 to 8.3 percent. The company describes this benchmark as a stress test rather than a measure of general error rates. GPT-6 Astra still holds the lead on this metric with a 5.5 percent error rate. Internal finance tests showed similar progress, with the model gaining 7.2 points on Financial Modeling and 11.6 points on Financial Research.

Pricing changes accompany the performance upgrades. API costs fall by 33 percent, with input priced at $2 per million tokens and output at $10. Cache hits cost $0.20 per million tokens. Although the new model generates 14 percent more output tokens, the total cost to run each Intelligence Index task is $1.98. This figure is nearly identical to the $2.00 cost for the previous version.

The model architecture remains consistent with the previous generation, featuring 2.8 trillion total parameters and 104 billion active parameters. It supports a one-million-token context window and accepts both text and image inputs. Artificial Analysis places the model on the Pareto frontier, indicating no comparable model achieves a higher score for its size. A faster variant, K3.1 Pro Highspeed, is also available and runs two to three times faster than the standard version.

Security considerations shaped the release strategy. Moonshot removed cyber-related data from both training phases for the open-weights version. On an internal benchmark reproducing recent software vulnerabilities, the uncleaned model solved 90.5 percent of tasks. After data cleaning, that figure dropped to 81.0 percent. DeepSeek-V4.1-Flash scored 85.7 percent, while the previous Kimi K3 model solved only 33.3 percent of the tasks.