Kimi K3.1 Pro Arrives on Vega
Kimi K3.1 Pro has arrived on vega.chat alongside its launch on Kimi's platforms. Moonshot AI's new "3T-class" flagship combines a 1M-token context window with native vision and stronger coding, 3D modeling, and finance capabilities. A full open-weight release is promised by November 4, 2026. Launching alongside it, K3.1 Pro Highspeed offers identical capabilities at roughly 2–3 times the speed for interactive use.
What is Kimi K3.1 Pro?
The first model in Moonshot AI's K3.1 series, K3.1 Pro is a post-trained version of Kimi K3 with significantly more post-training compute. "3T-class" describes its model class: Moonshot has not disclosed exact total or active parameter counts.
Its 1,000,000-token context window covers large codebases, lengthy financial documents, and extended agent sessions in a single request. Native visual understanding lets it work with screenshots, sketches, and reference images.
Day-one availability includes Kimi.ai and its mobile apps, Kimi Work, Kimi Code, and the Kimi API. Kimi Enterprise administrators can enable it for their organizations; access is off by default at launch.
Kimi API pricing: about a third below kimi-k3
Moonshot's published Kimi API pricing is:
| Token type | Price per 1M tokens |
|---|---|
| Input | $2.00 |
| Output | $10.00 |
| Cache write | $2.00 |
| Cache hit | $0.20 |
These are Kimi's vendor prices, not VEGA prices.
For comparison, kimi-k3 listed uncached input at $3.00, output at $15.00, and cache hits at $0.30 per million tokens. K3.1 Pro comes in about a third cheaper across those three categories, which makes previously cost-limited workloads worth another look.
Kimi K3.1 Pro benchmarks: where it improves
Agentic coding and terminal work
On Terminal-Bench 4.0, which evaluates realistic terminal tasks with strict tests, K3.1 Pro scores 65.7%, versus GPT-6 Astra's 58.2% and Claude Opus 5.5's 66.4%.
The comparison needs a caveat: K3.1 Pro used the Claude Code harness, while the other models ran at their maximum settings. These are not fully controlled, like-for-like results.
On Moonshot's in-house Kimi Code Bench 3.0 Long-horizon evaluation, covering multi-step coding and tool use, it reaches 61.2%, up from K3's 55.95%.
3D modeling, CAD, and 3D Make
Parametric 3D modeling is an unusual focus for a flagship release. From text, sketches, or reference images, K3.1 Pro infers shape, proportions, and spatial relationships. It writes and runs Blender/CAD scripts without MCP connectors, editing geometry, materials, and scenes.
BenchCAD results show both its strengths and remaining gaps:
| BenchCAD setting | K3.1 Pro | Comparison scores |
|---|---|---|
| With tools | 92.7 | Opus 5.5: 96.2; Astra: 95.9 |
| Without tools | 72.8 | Opus 5.5: 73.0; Fable 5.1: 60.6; DeepSeek-V4.1-Flash-1017: 41.2 |
The with-tools result uses Voxel IoU on a 1,000-file subsample.
Moonshot also ships a finished product around this: 3D Make is live on kimi.ai and in its app, with preview, editing, and export for 3D printing or as video and Live Photos. That product availability is specific to Kimi's platforms.
Finance and document analysis
K3.1 Pro scores 64.4 on OfficeQA Pro v2, which tests financial-document analysis, cross-verification, and multi-step computation. Moonshot claims this is state of the art. It also reports 58.6 on Frontier Finance, covering investment workflows from screening through profitability analysis.
Its in-house finance results improve over K3:
| Evaluation | K3.1 Pro | K3 |
|---|---|---|
| Financial Modeling | 57.6 | 50.4 |
| Financial Research | 70.4 | 58.8 |
| Business Operation | 37 | 8.9 |
Factual accuracy
On Moonshot's in-house Kimi Hallu Bench Hard, claim-level hallucinations fall from 23.4% for K3 to 8.3%. The share of answers containing at least one factual error drops to 42.4% of the predecessor's figure.
The with-tools chart reports these claim-level error rates, where lower is better:
| Model | Error rate |
|---|---|
| GPT-6 Astra | 5.5% |
| Kimi K3.1 Pro | 8.3% |
| Claude Opus 5.5 | 15.0% |
| GLM-5.3 | 22.6% |
| DeepSeek-V4.1 | 23.3% |
| Claude Fable 5.1 | 23.3% |
Moonshot explicitly describes this as a harsh stress test, not an estimate of everyday error rates.
Under the hood: joint RL and cyber safety
Moonshot attributes the gains to joint reinforcement learning across hundreds of downstream domains, supported by whole-stack optimization: RL infrastructure, long-context inference, training efficiency, sandboxes, rollout scheduling, and environment quality control.
The team also observed vulnerability-research capability increasing with post-training scale. For the open-weight release, it removed cyber-specific data from both mid-training and RL stages while retaining ordinary coding tasks, and conducted targeted evaluation using its in-house xdev-bench-hard benchmark.
Try Kimi K3.1 Pro on Vega
For developers, this is currently the strongest open-weight candidate for agentic coding with long context. For 3D teams, test shape inference and the generated CAD scripts; for finance teams, test document cross-checking and multi-step calculations. Kimi's lower vendor pricing also makes this a useful moment to revisit budget-sensitive workloads.
Several results are company-reported or internal, harness conditions differ, and parameter counts remain undisclosed. Pilot on your own tasks before committing.
Kimi K3.1 Pro is available in the vega.chat model picker, and developers can also call it through api.vega.chat, the OpenRouter-compatible API.