Moonshot's Kimi K3.1 Pro Lands on Microsoft Foundry With a 1 Million-Token Context at $11.50 per Million Output Tokens

Microsoft has added Moonshot AI's Kimi K3.1 Pro to the Foundry model catalog, making the Chinese lab's open-weight model available to enterprise developers on Microsoft's cloud platform as of November 5. The listing was published on the Microsoft tech community blog and picked up by Chinese financial press within hours.

Kimi K3.1 Pro is an open-weight model built for agentic work, accepting both text and images, with a context window of up to one million tokens. Microsoft lists pricing at $2.30 per million input tokens, $0.23 per million cached tokens and $11.50 per million output tokens. The company points to coding and research as the natural fits: agents that investigate a change across a large repository, or that need to read lengthy reports and supporting documents before producing a synthesis.

According to China's STCN Daily, the listing is notable for what it says about the business arrangement rather than the model itself: Kimi K3.1 Pro is the first revenue-share open-weight model on Microsoft's cloud, and Microsoft is Moonshot's second global hyperscaler partner after Amazon Web Services. Moonshot previously open-sourced the K3.1 family, with the Pro variant reported at around three trillion parameters, and has been pushing the series into enterprise channels aggressively this quarter.

Microsoft's guidance to developers is unusually candid about the limits of context windows. The announcement recommends judging completed tasks rather than token capacity alone, starting from a real workflow with the incumbent model as a baseline, and comparing finished work quality, long-input reliability, tool use, visual handling, speed and cost before switching. Microsoft attributes strong coding, agentic and multimodal evaluation results to Moonshot but publishes no numerical benchmark scores of its own, so those strengths remain vendor-reported.

The distribution race among Chinese AI labs into Western enterprise clouds is intensifying. Being listed on a hyperscaler catalog is now table stakes; the interesting shift is revenue-share arrangements, which give the platform an incentive to push the model and give the lab a recurring revenue channel that does not depend on its own international API sales.

For buyers, the practical takeaway is that a one-million-token window costs $2.30 per million input tokens, which undercuts most frontier proprietary models on raw price, but Microsoft itself cautions that stuffing more material into the prompt is no guarantee of better finished work. The honest test is task-level: feed the same real workload to K3.1 Pro and to the model already in production, and compare what comes out the other end, including what it costs.