Moonshot AI has released Kimi K3.1 Pro, a reasoning model that ranks fourth on the Artificial Analysis Intelligence Index, while reducing standard API token prices by one-third compared with its predecessor. The company plans to publish the model’s weights by November 4, 2026.
K3.1 Pro scores 55 on the index, behind Claude Opus 5.5 at 58 and Claude Sonnet 5.5 and Muse Wonder Full at 56 each. Artificial Analysis measures an 11-point improvement over Kimi K3. If the planned weights release proceeds and the rankings remain unchanged, K3.1 Pro would be the highest-scoring model with publicly available weights on that index.
The release is an expanded post-training effort rather than a new model architecture. K3.1 Pro retains K3’s mixture-of-experts design, with 2.8 trillion total parameters and approximately 104 billion active for each token. It supports text and image inputs and a context window of about one million tokens. Mixture-of-experts systems activate only part of their parameter pool for each token, reducing computation relative to activating the entire network.
The strongest reported gains include tasks in which an AI agent uses tools to complete work. Artificial Analysis assigns K3.1 Pro an Elo rating of 1,846 on GDPval-AA, up 312 points from K3 and close to Claude Opus 5.5 at 1,866 and Muse Wonder Full at 1,857. Such ratings describe relative performance within an evaluation rather than a percentage of tasks completed correctly.
Moonshot’s own testing puts K3.1 Pro at 65.7% on Terminal-Bench 4.0, which evaluates command-line tasks. Its comparison gives Claude Opus 5.5 a score of 66.4% and GPT-6 Astra 58.2%. These vendor-published results are separate from the Artificial Analysis assessments and should not be treated as independently replicated measurements.
The company also reports improved reliability on its internal Kimi Hallu Bench Hard evaluation. The claim-level hallucination rate fell from 23.4% for K3 to 8.3% for K3.1 Pro. Moonshot describes this as a stress test, not an estimate of how often the model will make errors in ordinary use. GPT-6 Astra recorded a lower rate of 5.5% in the same comparison.
For developers, standard API pricing is $2 per million input tokens and $10 per million output tokens, down from $3 and $15 respectively. Cached input costs $0.20 per million tokens. Lower token prices do not necessarily translate into an equivalent reduction in task costs: Artificial Analysis measured spending of $1.98 per Intelligence Index task, versus $2.00 for K3, with the newer model producing more output tokens.
A separate K3.1 Pro Highspeed variant is also available. Moonshot says it runs two to three times faster, with API prices of $4 per million input tokens and $20 per million output tokens. Separate benchmark results for that version have not been published, leaving its performance relative to the standard model unclear.
Moonshot has also detailed changes intended to limit offensive cybersecurity capabilities ahead of the weights release. Removing cyber-related training data reduced performance on its internal vulnerability-research evaluation from 90.5% for an unrestricted checkpoint to 81.0% for the release version. The evaluation includes tasks involving recently disclosed software vulnerabilities.
Both variants are accessible through Kimi’s consumer services, Kimi Work, Kimi Code and the API. The planned November weights release would allow deployment outside Moonshot’s hosted services, subject to hardware requirements and licensing terms that have not yet been specified.
Sources: The Decoder, Pasquale Pillitteri, OrcaRouter, OfficeChai