Kimi K3.1 Pro: access, pricing and what changed

Moonshot lists Kimi’s website, Kimi Work, Kimi Code and the Kimi API as access routes. Start with the official website for an interactive product, use Kimi Code for supported coding workflows, or choose the API when building your own application. Each route has its own account, usage limits and billing rules; a subscription to one product should not be treated as credit for another.

For the China API platform, the official guide says a completed top-up is required to unlock this flagship model, with a minimum top-up of RMB 10. Its new-user vouchers cannot be used for K3.1 Pro. Check the rules for the platform you actually use before assuming that an account or promotional balance includes access.

The two K3.1 variants share the advertised intelligence and one-million-token context window. Moonshot describes Highspeed as roughly two to three times faster. Pro is the lower-priced choice in the current international price list; Highspeed is an option when faster responses matter enough to justify its higher token price. Actual response time depends on the request, provider and reasoning workload.

Compared with K3, the K3.1 announcement emphasizes additional post-training and stronger coding, 3D and financial workflows. Its headline improvement is not a larger context window: both generations advertise 1M tokens. Official benchmark results are useful context, but they are not measurements of this site or a guarantee that every task will improve. Test representative prompts with a fixed budget and the same evaluation criteria.

For coding, useful evaluations include fixing a failing test, modifying a small real repository and reviewing the resulting diff. Ask for assumptions, verification commands and the smallest necessary change. For 3D work, the model can help write scripts for Blender or CAD tools, but generating text is different from executing those tools and exporting an asset. Your environment must provide the required software, permissions and validation.

For document or financial analysis, supply dated source material, units and explicit calculation rules. Require citations back to the supplied evidence and separate missing information from estimates. A model’s financial benchmark score does not replace checking formulas or source documents. The examples below are reusable task briefs, not recorded K3.1 outputs or evidence that this site runs an agent, a CAD tool or a spreadsheet engine.

The official model IDs are kimi-k3.1-pro and kimi-k3.1-pro-highspeed. K3.1 Pro always uses thinking; reasoning_effort supports low, high and max, with max as the documented API default. Review any adapter that disables thinking, assumes immediate answer tokens or keeps only the assistant’s final text. The official multi-turn contract requires the full assistant message to be returned in conversation history.

Audit streaming, output limits, cancellation, retries and billing before changing production traffic. Reasoning and final answers arrive in separate fields. Budget for the complete request, not only the visible answer, and check current parameter restrictions in the official quickstart. The provider’s 1M context capacity also does not override smaller limits imposed by your own application.

As of this guide’s update, Moonshot says the complete model weights will be released by November 4, 2026. Hosted availability does not mean that a downloadable checkpoint, final license or compatible local package is already available. Recheck the official release before choosing a serving stack or planning a migration around self-hosting.

A 3T-class model is not automatically a consumer-laptop deployment. Memory, quantization, serving support and hardware requirements depend on the actual checkpoint and runtime. We do not provide an unverified Ollama command or promise that the full model fits on a consumer GPU. For now, use the official hosted access routes if you need to evaluate it.

The official API pricing notice says K2 models stop opening to newly registered API users on October 31, 2026, and that all K2 models are planned for retirement by the end of December. This refers to upstream API accounts, not to new registrations on an independent application. Existing access on October 31 does not remove the need to prepare for the later retirement.

For an existing app, first verify account access and make a separate candidate integration. Compare representative single-turn and multi-turn tasks, interrupted streams, error handling and cost. Keep the selected model visible to users and retain a rollback plan while the old API remains available. Chat on this site lets you select K2.6 or K3.1 Pro, with application-level context and output limits. It does not provide the full official Kimi product, a 1M-token chat allowance or integrated CAD execution.