What is available today
Moonshot AI has made Kimi K3.1 Pro accessible via app and API, but as of October 28 the model's weights are not yet downloadable. The company has announced publication by November 4, 2026. Until that date, those who want to try the model depend on Moonshot's services and cannot run it on their own infrastructure or independently verify its behavior.
The distinction is relevant for businesses evaluating a model for products or internal processes. API access allows immediate testing of features, while the weights would enable execution on customer-controlled hardware, in compliance with the license and technical requirements. Moonshot presents K3.1 Pro as a 3T-class open-weight model with context up to one million tokens, but it places the full release of the weights at a future date. The model is already selectable on Kimi.ai, Kimi Work, Kimi Code and via API.
Prices and stated benchmarks
According to the price list published by Kimi, the standard API rate is lower than that of the previous K3 for ordinary tokens: $2 per million input tokens and $10 per million output tokens, versus $3 and $15. The price list distinguishes cache writing from cache reading: input tokens already stored cost less, while output continues to be counted. The effective cost per task therefore depends on the mix of input, generated responses and context reuse. For the Highspeed variant, Moonshot states a speed two or three times higher, with double rates for ordinary tokens.
On performance, Moonshot attributes to K3.1 Pro a 65.7% success rate on Terminal-Bench 4.0, which measures completion of agentic tasks in terminal environments, versus 58.2% for GPT-6 Astra and 66.4% for Claude Opus 5.5. The result concerns a system composed of model and software agent: tools, configurations and number of attempts affect the measurement. The same company notes that the table combines scores collected from different sources and that the benchmark was updated with new resources, tasks and verifiers; moreover, an independent review by Epoch AI reported scoring flaws in 30 of 66 tasks. The score should therefore be read as an initial indication, not as a sufficient measure for estimating enterprise productivity.
On BenchCAD, Moonshot reports 92.7% with tools and 72.8% without tools, on a subset of one thousand files. The test concerns 3D reconstruction via CAD code and volumetric similarity: it measures a specific capability, but does not by itself demonstrate that the model can replace design work or meet production constraints.
Security: claims awaiting verification
With the future publication of the weights, the model's behavior in dual-use cybersecurity activities also becomes relevant. Moonshot states that it removed vulnerability research data from two training stages; the internal comparison measures the effect of these interventions. In its xdev-bench-hard, the internal checkpoint without such interventions solves 90.5% of problems in three attempts, while the final model reaches 81%. This is a difference of 9.5 percentage points compared with the internal checkpoint, not an independent security certification. Moonshot also reports that the final model exploited vulnerabilities in an isolated environment, with active mitigations.
Until the weights and a reproducible protocol are available, external laboratories will not be able to verify these results under the same conditions. For those who require infrastructure control, audits and independent tests, the decisive step remains set for November 4: after publication, it will be possible to evaluate the license, hardware and behavior of the installed model.
