Kimi K3.1 Pro:
Frontier-Level, but Not Yet Open
Moonshot's new model brings long-context coding, agents, 3D/CAD work and a competitive independent intelligence score. The interesting part isn't only the benchmark numbers — it's what has been verified, what comes from Moonshot's own testing, and what hasn't shipped.
Launch at a glance
Released Oct 23–24 across time zones. As of Oct 24, the weights and release license are not publicly available. “Open-weight” describes the company's announced plan, not an already downloadable model.
01 / Independent intelligence result
Artificial Analysis' Intelligence Index (version shown Oct 24): K3.1 Pro scores 55 and ranks #4 among 215 listed models. This is a composite score, not a percentage and not proof it wins every coding task.
Kimi K3.1 Pro · AA Intelligence Index
Single-model score shown on a 0–100 plot; no invented comparison bars. AA also measured ~73.5 output tokens/sec for the standard API model, calling it slower and more verbose than average. Different providers or versions may vary.
02 / Coding: Terminal-Bench 4.0
Moonshot's published comparison for practical terminal-agent tasks. Each source/harness may differ. The Kimi result uses the Claude Code harness. Vendor-assembled comparison — not one independently controlled head-to-head run.
Success rate · higher is better
Source: Moonshot K3.1 tech blog, published Oct 24, 2026. The 65.7% figure is distinct from Vals' independent 63.64% measurement due to differing setups.
Kimi Code Bench 3.0 Long-horizon · Moonshot internal
The internal coding benchmark covers multi-step execution, tool use and end-to-end engineering tasks. It has not been independently replicated here.
03 / 3D and CAD: BenchCAD
Moonshot's with-tools results on a randomly sampled 1,000-case subset of Vision2Code. Metric: voxel intersection-over-union. Strong, but different from shipping reliable CAD designs without human review.
BenchCAD · with tools · higher is better
Moonshot reports end-to-end game and web development and support for Blender, CAD, Unity, Unreal and Godot. These are announced capabilities and examples, not a substitute for independent production testing.
04 / Hallucinations: the precise claim
Moonshot's internal Kimi Hallu Bench Hard uses deliberately difficult factual prompts. A lower claim-level hallucination rate is better.
Claim-level error rate · lower is better
05 / API cost
Published API rates in USD per million tokens. The comparison below uses Kimi K3 and K3.1 Pro. Highspeed costs more; its faster output is a vendor claim.
Input tokens · lower is better
Output tokens · lower is better
| Model | Input / 1M | Output / 1M |
|---|---|---|
| Kimi K3 | $3.00 | $15.00 |
| Kimi K3.1 Pro | $2.00 | $10.00 |
| K3.1 Pro Highspeed | $4.00 | $20.00 |
Cache-hit input is listed at $0.20 / 1M for K3.1 Pro. Rates reflect Oct 24 listings; verify current pricing before purchasing. Lower price per token doesn't always mean lower cost per completed task: verbosity, retries and tool usage also matter.
Open weights and the security trade-off
Moonshot says it intends to publish the full model weights by November 4, 2026. At the time of this report, the company has not shipped those weights or named the license. Until then, you can use the hosted tools and API, but should not assume you can self-host K3.1 Pro.
The company also says it deliberately filtered cyber-related material during training to reduce misuse risk. On its internal vulnerability-exploitation benchmark, the unfiltered checkpoint reached 90.5% while the released K3.1 Pro configuration scored 81.0%. Reuters independently reported the company's decision, but not an independent replication of those scores.
Where developers can try it
Use Kimi, Kimi Work, Kimi Code (switch with /model), or the Moonshot API with IDs kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Windows users can use the Kimi Work desktop app version 3.1.0 or later.
What I would actually test
Try the same multi-file coding issue, browser/tool workflow and small CAD-to-geometry task on K3.1 Pro and your current agent model. Track completed tasks, tool-call failures, turnaround time, generated tokens and final API bill. Treat website demos and vendor leaderboards as reasons to test — not as results from your own workflow.
Sources and verification
- Moonshot AI — official K3.1 Pro technical blog (Oct 24, 2026) · specification, company claims, methodology, coding/CAD and internal reliability tests
- Artificial Analysis — independent intelligence, speed and API pricing
- Vals AI — independent index and terminal-agent results
- Reuters — cyber-capability controls and promised weight release (Oct 24, 2026)
- Vercel AI Gateway — provider availability and pricing