Have a project in mind? Let's talk.
kimi k3.1 pro arrived from Moonshot AI on November 2, 2026, and the most important line in its tech blog is not a benchmark score. It is the admission that Moonshot found cyber-related content inside its general training data, removed it from both the mid-training corpus and the reinforcement learning environment mix, and shipped an open-weight model whose offensive cyber capability is deliberately below what its own pipeline could produce. That decision sits inside an otherwise conventional frontier-model release: a 3T-class open-weight model with a 1-million-token context window, a Highspeed variant, API pricing at $2 per million input tokens, and full weights promised by November 4. This post works through Moonshot's tech blog, API documentation and pricing page to cover what the model is, how the benchmark claims should be read, the cybersecurity decision in detail, and what it means for teams considering self-hosted open weights. Source
Per Moonshot's tech blog, K3.1 Pro is a post-trained version of Kimi K3 with substantially larger post-training compute, produced by scaling joint reinforcement learning across hundreds of downstream domains. Moonshot calls it its most capable 3T-class open-weight model to date, with a 1-million-token context window. Availability is broad on day one: Kimi.ai and the mobile apps, the Kimi Work desktop app, Kimi Code via the /model command, and the Kimi API under the model names kimi-k3.1-pro and kimi-k3.1-pro-highspeed, with Zero Data Retention available for API customers. Kimi Enterprise administrators can enable it for their organization, and access is off by default at launch. Full model weights are promised by November 4, 2026, so at publication time the "open-weight" label is a commitment, not yet a downloadable artifact. Source
Moonshot reports K3.1 Pro at 65.7% on Terminal-Bench 4.0, ahead of GPT-6 Astra at 58.2% and narrowly behind Claude Opus 5.5 at 66.4%, and 92.7% with tools on a 1,000-file BenchCAD subset against 96.2% for Opus 5.5. Read these with the footnotes attached. K3.1 Pro was evaluated with the Claude Code harness on several coding benchmarks while some competitor scores come from vendor blogs or official leaderboards, so harness conditions differ across the column. Other headline numbers come from Moonshot's own internal benchmarks: Kimi Code Bench 3.0 Long-horizon at 61.2% versus K3's 55.95%, and a claim-level hallucination rate on its internal Kimi Hallu Bench falling from 23.4% for K3 to 8.3%. Internal benchmarks are useful for tracking a vendor's own progress, but they are not independently verifiable, and we did not find independent replication of the headline scores at publication time. Treat all of it as Moonshot's evaluation of its own model until third parties publish matched-harness results. Source
This is the section that matters for security teams. Moonshot says that as post-training scaled, vulnerability research performance improved steadily, and tracing the cause revealed cyber-related content embedded in the general mid-training corpus plus cyber-related tasks in the RL environment mix. For the open-weight release, it removed that data from both stages while keeping general-purpose coding tasks. On Moonshot's in-house xdev-bench-hard, which asks an agent to develop working exploits for faithfully reproduced, recently disclosed vulnerabilities in software like the Linux kernel, Node.js and Exim, the checkpoint without these controls solves 90.5% of problems at pass@3 while the shipped K3.1 Pro reaches 81.0%. Moonshot frames its release principles explicitly: do not release open weights that extend the absolute frontier of cyber capability, keep the model behind frontier proprietary systems on offensive tasks, and progressively raise the cyber capability available in open weights so defenders can audit code they maintain, including legacy systems that rarely get security reviews. Note the tension it leaves open: 81.0% on its own exploit benchmark is still a high absolute number, and the evaluation is internal, so the practical residual capability is something defenders should test rather than assume. Source
Moonshot's pricing page lists K3.1 Pro at $2.00 per million regular input tokens, $10.00 per million output tokens and $0.20 per million cached input tokens, with cache writes at $2.50 for the default 5-minute TTL or $4.00 for a 1-hour TTL. The Highspeed variant, which Moonshot says delivers two to three times the speed, costs $4.00 input and $20.00 output. Pricing is flat across the context window with no long-context tier, a contrast with the above-272K surcharge we covered in our gpt-6.1 terra post. The platform documentation adds that kimi-k3.1-pro supports batch inference at 60% of the standard price for non-real-time work, and that the K2 series is closed to new registrations as of October 31, 2026 with full retirement planned for the end of December 2026, so existing K2 users should plan their migration. Source
One operational detail deserves its own paragraph because it is easy to miss in the pricing documentation: the K2 series closed to newly registered users on October 31, 2026, and Moonshot plans to retire all K2 series models by the end of December 2026. K2 models are limited to a 262,144-token context window, while all K3 series models support 1,048,576 tokens, so the migration is also a context upgrade. Teams with production workloads pinned to K2 model names have roughly two months to test K3 or K3.1 Pro equivalents, update model identifiers in their configurations, and re-run their evaluations, since pricing, context behavior and capability profiles all change between series. If your agent stack abstracts the model behind configuration, this is a planned afternoon of work; if model names are hard-coded across services, it is a small incident waiting for a calendar date. Source
This section is our read, separate from the sources. Three releases in the past week tell one story: longcat-2.5 from Meituan, GPT-6.1 Terra through Bedrock, and now K3.1 Pro. Open-weight models are close enough to proprietary frontier scores on vendor-run benchmarks that the build-versus-buy question is now about governance, not just capability. If you self-host K3.1 Pro after the weights drop, you inherit responsibility for the controls Moonshot applied at training time and for everything it could not control: prompt-level misuse, tool access, and what your agents are allowed to do with the model's output. Our OWASP agentic AI top 10 breakdown maps those runtime risks; training-time data removal does nothing about them. On the security-specific claim, Moonshot's argument that gradually diffusing dual-use capability makes software more secure is a position, not a finding, and the reasonable operational stance is to evaluate the released weights on your own defensive tasks before granting them tool access in production. The flat pricing and 1M context make K3.1 Pro genuinely interesting for long-document code auditing, which is probably the use case where its design choices align best with defensive work. Source
kimi k3.1 pro is Moonshot AI's newest open-weight model, released November 2, 2026: a post-trained version of Kimi K3 with a 1-million-token context window, available on Kimi.ai, Kimi Work, Kimi Code and the Kimi API, with full weights promised by November 4, 2026. Source
$2.00 per million regular input tokens, $10.00 per million output tokens and $0.20 per million cached input tokens, with cache writes at $2.50 for 5-minute TTL or $4.00 for 1-hour TTL. The Highspeed variant costs $4.00 input and $20.00 output, and batch inference bills at 60% of standard pricing. Source
Moonshot says it found cyber-related content in its general training data and RL mix, and removed it for the open-weight release so the model's offensive cyber capability stays below what its pipeline could produce and behind frontier proprietary models. On its internal xdev-bench-hard exploit benchmark, the unmodified checkpoint scores 90.5% at pass@3 versus 81.0% for the shipped model. Source
On Moonshot's reported numbers, K3.1 Pro scores 65.7% on Terminal-Bench 4.0 versus 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra. These figures come from Moonshot's own evaluations with differing harnesses across models, and we did not find independent replication, so treat them as vendor claims pending third-party testing. Source
Moonshot says full model weights will be released by November 4, 2026. Until then, access is through Kimi's hosted products and the Kimi API, where Zero Data Retention is available for API customers. Source
Ready to take the first step towards unlocking opportunities, realizing goals, and embracing innovation? We're here and eager to connect.