Microsoft adds Kimi K3.1 Pro to Foundry with a million-token context window

Developers using Microsoft Foundry now have another option for AI agents that work through large codebases and lengthy documents. Microsoft added Moonshot AI’s Kimi K3.1 Pro to its model catalog on November 5, 2026. The open-weight model accepts text and images, with room for up to one million tokens of context.

The announcement concerns availability through Foundry. Microsoft describes K3.1 Pro as the first model in Moonshot’s K3.1 series, building on Kimi K3 with improvements in coding, research and other tasks involving several steps. Those improvements are Microsoft’s description of the model, rather than a new head-to-head evaluation published with the Foundry announcement.

More source material in the same task

A context window is the amount of material a model can work with in a request. Microsoft positions K3.1 Pro’s large window as useful for bringing more relevant material into an agent’s work. Its image understanding adds another kind of source: charts, images and document pages can sit alongside text in a broader task.

For coding, Microsoft points to agents investigating a change across a large repository, understanding related files and working through an engineering task. For research, it describes supplying lengthy reports and supporting documents so the model can examine source material before producing a synthesis. These are proposed application uses, not measured outcomes from customer deployments.

The distinction is between giving an agent enough material and getting useful work back. Microsoft explicitly cautions that a large context window is not a shortcut to a good application. The model still needs to use that information effectively, finish the task, and meet the customer’s requirements for cost and speed.

$2.30Input

Microsoft lists input pricing at $2.30 per million tokens.

$0.23Cache

Microsoft lists a cache price of $0.23 per million tokens.

$11.50Output

Microsoft lists output pricing at $11.50 per million tokens.

Token rates versus the cost of finishing

The pricing table separates input, cache and output charges. Microsoft’s evaluation guidance uses a different measure: cost per successful task. It recommends comparing the quality of completed work, not just the first response, and checking whether the model handles long inputs reliably. That puts the published rates alongside an application-level test, rather than treating price alone as the selection rule.

Benchmarks are the starting point

Microsoft says Moonshot reports strong results across coding, agentic and multimodal evaluations—tests involving software work, multi-step agents, and combinations of input types. Its announcement does not provide numerical benchmark results or a scored comparison with named rivals. Instead, Microsoft argues that those results cannot settle whether the model will perform well in a particular application.

For developers deciding whether to switch, Microsoft proposes a workload-based comparison:

• Start with a real coding, research or document workflow and establish a baseline using the model already in place.
• Check the finished work, including reliability with long inputs, use of tools and performance on visual material.
• Compare K3.1 Pro with other models in Foundry, rather than assuming one model fits every step.