Moonshot launches Kimi K3.1 Pro, with open weights promised by November 4

AI models · October 23, 2026

Moonshot AI has released Kimi K3.1 Pro, giving developers another large-model option for coding and extended work with documents, tools and visual inputs. The company’s October 23 release notes confirm that the model is live in Kimi Code, alongside a faster Highspeed variant.

The launch comes with an important distinction: hosted access is available now, while Moonshot says the full model weights will be released by November 4. Calling it an open-weight model does not mean developers can already download and deploy this version themselves. Moonshot’s launch announcement

A K3 upgrade aimed at longer jobs

Moonshot describes K3.1 Pro as a post-trained version of K3, using more post-training computation. Its stated priorities include coding, 3D creation and financial research. Technical announcement

The official API platform lists the new model as a three-trillion-parameter-class system with native visual understanding and a one-million-token context window. A large context window can accommodate more material in one request; it does not, by itself, establish that every detail will be recalled or used correctly.

Promising benchmark claims, not a blanket lead

In Moonshot’s reported Terminal-Bench 4.0 comparison, K3.1 Pro scores 65.7%, ahead of the cited GPT-6 Astra result of 58.2% and behind Claude Opus 5.5 at 66.4%. Moonshot says it tested Kimi with the Claude Code harness and drew rival scores from release reports or the benchmark leaderboard. These are company-reported comparisons, not independent testing by PromptsZone. Results and methodology

The practical question is whether those gains survive a team’s own codebase, tool configuration and review process. A benchmark result is a reason to evaluate a model, not a guarantee that it will finish a particular project.

Lower standard API prices, a premium for speed

Moonshot’s API platform lists the following US-dollar rates per million tokens:

• K3.1 Pro: $2 for input and $10 for output.
• K3.1 Pro Highspeed: $4 for input and $20 for output.

The standard rates are one-third below the listed K3 rates of $3 for input and $15 for output. That is a reduction in token prices, not a demonstrated one-third saving per completed task.

The billing documentation says API usage is pay-as-you-go, separate from Kimi memberships, and prices exclude applicable taxes. Cache billing also affects the total. Teams comparing costs should measure an entire completed workflow, including retries, tool use and any additional billable services.

What developers can use today

According to Kimi Code’s model documentation, standard K3.1 Pro is available on Plus and higher plans; the full 1M context and Highspeed option require Pro or higher. Legacy-plan names differ. Highspeed is advertised at roughly two to three times the output speed, with about twice the quota consumption.

The model requires thinking mode. Moonshot recommends starting a fresh session instead of switching an existing conversation to K3.1 Pro. It also warns that a mistyped Kimi Code model ID can fall back to a different model, so developers should verify what their client actually selects. Configuration and access requirements

For the separate pay-as-you-go service, the API quickstart names the models kimi-k3.1-pro and kimi-k3.1-pro-highspeed and says access requires a successful top-up of at least $1. Its identifiers should not be confused with the shorter Kimi Code model IDs.

The next test is deployment

Hosted evaluation can begin now. Self-hosting decisions should wait for the actual K3.1 Pro weights and their accompanying license. This report has not verified a downloadable K3.1 Pro release or final licensing terms, and the previous K3 license should not be assumed to carry over.

For developers, the strongest reason to pay attention is the combination of a new capability claim, lower standard token prices and a choice between speed and cost. The useful next evidence will be reproducible workload results and the promised weight release, rather than broader claims of model supremacy.

Sources and reporting note

This article uses Moonshot’s launch announcement, product documentation and published API rates, checked on October 23, 2026. PromptsZone has not independently benchmarked the model. Launch-day documentation can change; verify access, billing and license terms before deployment.

Illustrative archival photo: Lynn Freeny / U.S. Department of Energy, via Wikimedia Commons. Summit at Oak Ridge National Laboratory, October 18, 2017; not a Moonshot AI facility. Public domain (U.S. government work). Display resized/cropped.