Moonshot launches Kimi K3.1 Pro for coding and research, with model weights due in November

Moonshot AI has launched Kimi K3.1 Pro across its hosted products and developer API, targeting longer coding jobs and financial research as competition among advanced AI models intensifies. The company says downloadable model weights will follow by November 4. Customers can use the service now, while organisations planning to run the model on their own infrastructure still face a separate release milestone.

The model follows Kimi K3 and supports a context window of about one million tokens. Moonshot also offers a Highspeed version, which it says runs two to three times faster. The company's launch materials describe improvements in coding, three-dimensional design and research involving financial documents. Those are the capabilities it wants buyers to evaluate, rather than evidence that every workflow will become faster or more profitable.

An available API and downloadable weights create different options

Hosted access lets a developer send requests to a provider's infrastructure and pay for processing. Downloadable weights would allow a different deployment model, subject to the eventual release terms and the computing resources required. The promised November date should therefore remain distinct from the October 23 service launch. Calling the weights already available would overstate what Moonshot has delivered so far.

This matters to procurement teams comparing operating cost, control and implementation time. Using a hosted model removes some infrastructure work but adds a provider relationship and usage charges. Operating a model independently moves more responsibility to the customer. The economics of running AI include utilisation, technical support and the resources needed to deliver an acceptable result, as well as the model's nominal price.

Kimi's documentation supports tool calls and workflows in which the model interacts with software rather than only returning a chat response. A coding agent might inspect information, propose changes and invoke tools over several steps. Each step adds opportunities to gather evidence, but also creates dependencies on the tools' permissions, outputs and error handling. Better reasoning does not remove the need to check the resulting work.

Price per token is a starting point for a cost comparison

OpenRouter lists the standard Moonshot endpoint at $2 per million input tokens and $10 per million output tokens, with cached-input pricing listed separately. Its Highspeed listing is $4 and $20 respectively, with the same listed cache-read rate. These are the provider listings checked for this report, not a guarantee that every channel, plan or future request will carry identical charges.

At the standard listed rates, one million uncached input tokens and 100,000 output tokens would cost $3 before other applicable charges. That is a hypothetical calculation: $2 for the input plus $1 for the output. Repeated attempts, additional tool use and larger outputs can change the bill. A lower quoted rate only becomes a meaningful saving if the system completes the required task at an acceptable quality.

The official pricing explanation distinguishes API usage from subscriptions to Kimi's other products. It also describes billing for repeated context and separately priced cache behaviour. A long context window establishes how much information the system can accept; it does not establish that every document will be relevant or that every important detail will be recalled correctly. Keeping the input focused can matter to both cost and reviewability.

Research claims still need to survive a real workflow

Moonshot's launch report presents stronger results on coding and financial-research evaluations. An organisation considering the system can test those claims against its own work: whether a calculation traces back to the correct filing, whether linked spreadsheet formulas remain coherent and whether generated code passes meaningful checks. The return on an enterprise AI deployment depends on completed, usable work after review and correction.

The company also describes deliberately reducing offensive cyber capability before the planned weight release by removing cyber-related training material. On its internal vulnerability benchmark, the final model scored 81%, against 90.5% for a comparison model without those controls. That is a 9.5-percentage-point gap on the company's test, not an independently verified measure of the chance that any particular system can be attacked.

The result illustrates the difficulty of separating useful coding ability from capabilities that can be misused. It does not establish that a model needs no operational controls. Access to sensitive information and tools remains a distinct deployment decision, as our discussion of enterprise AI and data security explains. A benchmark about code execution cannot substitute for a company's own access policy.

Kimi K3.1 Pro adds another product for buyers to compare on difficult coding and research work. The immediate tests are its output quality, completion time and total cost on those tasks. November's promised weight release will then determine which additional deployment options become available, and under what terms. Those two stages should be assessed separately as the model moves from a launch claim to repeatable use.