Moonshot cuts AI costs 33 percent
New Kimi K3.1 Pro model scores 55 on intelligence index, closing the gap with top proprietary systems.
By Thiago Kovalenko ·
The gap between open and proprietary artificial intelligence models is narrowing as Chinese firm Moonshot AI releases Kimi K3.1 Pro. The new reasoning model ranks fourth on the Artificial Analysis Intelligence Index, scoring 55 points.
This result represents an 11 point jump over its predecessor, Kimi K3. Only three systems currently outperform it: Anthropic's Claude Opus 5.5 at 58 points, Claude Sonnet 5.5 at 56, and Muse Wonder Full at 56. Moonshot plans to publish the model weights by November 4, 2026.
If the release proceeds as scheduled, K3.1 Pro will become the highest scoring open weights model available. That change would reduce the performance difference between open and proprietary systems from 12 points to 3. The company also offers K3.1 Pro Highspeed, a variant that runs two to three times faster than the standard version.
Performance gains are most visible in agentic tasks. K3.1 Pro reached 1,846 Elo on the GDPval-AA benchmark, a 312 point increase over K3. On Terminal-Bench 4.0, the model achieved a 65.7 percent score, surpassing GPT-6 Astra at 58.2 percent. The previous version, K3, managed only 13 percent on that specific test.
Moonshot also reports a sharp decline in hallucinations. On its internal Kimi Hallu Bench Hard, claim level hallucinations dropped from 23.4 percent with K3 to 8.3 percent. The company notes that GPT-6 Astra still leads this metric at 5.5 percent. Financial benchmarks showed similar improvements, with gains of 7.2 points in modeling and 11.6 points in research.
The model retains the same architecture as K3, featuring 2.8 trillion total parameters with 104 billion active. It supports a one million token context window. API prices fall by 33 percent, setting input costs at $2 per million tokens and output at $10. Cache hits cost $0.20 per million tokens.
Moonshot addressed cybersecurity in its technical documentation, noting that vulnerability research capabilities improved as post training scaled. For the open weights release, the company removed cyber related data from training phases. The uncleaned model solved 90.5 percent of tasks on the xdev bench hard, while the cleaned version solved 81.0 percent. Weights are expected to go public by November 4, 2026.
