Moonshot AI releases Kimi K3.1 Pro with lower costs and higher scores
The new model improves on agentic tasks and coding while cutting API prices by a third, challenging top-tier proprietary rivals.
By Petra Murphy ·
Moonshot AI has introduced Kimi K3.1 Pro, a post-trained update to its previous model that aims to close the gap with leading proprietary systems. The release marks a significant shift in the competitive landscape, as the new model climbs to fourth place on the Artificial Analysis Intelligence Index. It achieves a score of 55 points, placing it just behind Anthropic's Claude Opus 5.5, which holds the top spot with 58 points. This narrow margin suggests that Chinese developers are rapidly narrowing the distance between open and closed AI architectures.
The most notable improvements in K3.1 Pro appear in agentic tasks, where the model demonstrates substantial growth. On the GDPval-AA benchmark, the system reaches an Elo rating of 1,846, a jump of 312 points from its predecessor. This performance places it slightly behind Claude Opus 5.5 and Muse Wonder Full. In coding evaluations, the model reports a 65.7 percent success rate on Terminal-Bench 4.0. This figure exceeds GPT-6 Astra and trails Claude Opus 5.5 by a mere 0.7 percent. The predecessor, K3, had managed only 13 percent on the same test, highlighting the scale of the upgrade.
Beyond raw performance, Moonshot has adjusted the economic structure of the model. While the parameter count remains at 2.8 trillion total and 104 billion active, the company has reduced API prices by 33 percent. Input costs now stand at $2 per million tokens, down from $3, while output costs drop to $10 from $15. Cache hits are priced at $0.20. Despite these lower rates, the cost per Intelligence Index task remains nearly identical to the previous version at $1.98. This stability is attributed to the model generating 14 percent more output tokens, which offsets the per-token price reduction.
The company also emphasizes advancements in 3D modeling and specific industry applications. K3.1 Pro can interact directly with game engines such as Unity, Unreal Engine, and Godot. On the BenchCAD benchmark, it scores 92.7 percent when using tools, a result close to top competitors. Without tools, the model achieves 72.8 percent, nearly matching Claude Opus 5.5. Moonshot states that the system is particularly effective at parametric modeling of mechanical and architectural parts. A new feature called 3D Make allows users to preview, edit, and export 3D models directly on the platform.
Contextualizing these gains, the release is part of a broader trend where open-weights models challenge proprietary leaders. If Moonshot releases the model weights by November 4, 2026, as planned, K3.1 Pro would become the highest-scoring open-weights model available. This would shrink the gap between the best open and proprietary systems from 12 points to just 3. The company has also addressed cybersecurity concerns by removing cyber-related data from the training phases for the open release. This measure reduces the model's offensive capabilities on internal tests from 90.5 percent to 81.0 percent, aiming to balance utility with security while still supporting defensive auditing tasks.
