Kimi K3.1 Pro is Moonshot AI's new flagship model, and it scores 65.7% on Terminal-Bench 4.0 against 66.4% for Claude Opus 5.5. That is less than a point behind, at $2 per million input tokens and $10 per million output, with open weights due by November 4.
📺 Watch: Kimi K3.1 Pro is INSANE!
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside
Moonshot calls it its most capable 3T-class open-weight model so far. It has a 1M-token context window, and you can use it today in the Kimi app, Kimi Code and the Kimi API.
I covered it on my channel this week and ran the same prompts on Kimi and Claude side by side. Here is what the model is, how the benchmarks stack up, what it costs, where it loses, and how to start using it.
If you want the full Agent OS, a 30-day AI agent roadmap and daily tutorials as new models like this drop, the AI Profit Boardroom is $69/mo with four coaching calls a week.
What Kimi K3.1 Pro actually is
K3.1 Pro is the first model in Moonshot's K3.1 series. Moonshot describes it as a post-trained version of Kimi K3, built with a lot more post-training compute than K3 got.
So it is not a brand new base model. It is K3, trained harder on coding, knowledge work and agent tasks.
There are two versions:
- kimi-k3.1-pro: the standard model.
- kimi-k3.1-pro-highspeed: the same model, which Moonshot says runs 2 to 3 times faster, at a higher price.
Both have a 1M-token context window. Thinking is always on. You cannot switch the chain of thought off, but you can turn the reasoning effort down to low, high or max. Max is the default and the setting Moonshot used for its benchmarks.
Kimi K3.1 Pro benchmarks vs Opus 5.5 and GPT-6 Astra
Here are the headline numbers from Moonshot's own K3.1 Pro tech blog. These are Moonshot's tests, so read them as vendor claims.
| Benchmark (Moonshot's figures) | Kimi K3.1 Pro | Claude Opus 5.5 | GPT-6 Astra |
|---|---|---|---|
| Terminal-Bench 4.0 (higher is better) | 65.7% | 66.4% | 58.2% |
| BenchCAD 3D, with tools (higher is better) | 92.7 | 96.2 | 95.9 |
| BenchCAD 3D, no tools (higher is better) | 72.8 | 73.0 | not reported |
| Kimi Hallu Bench hard, with tools (lower is better) | 8.3 | 15.0 | 5.5 |
| Kimi Hallu Bench, no tools (lower is better) | 5.4 | 6.8 | 4.2 |
Three things jump out.
Coding is basically level with Opus 5.5. 65.7% against 66.4% on Terminal-Bench 4.0 is a rounding-error gap, and it is well clear of GPT-6 Astra's 58.2%. On Moonshot's internal long-horizon coding test, Kimi Code Bench 3.0, K3.1 Pro scores 61.2 against 55.95 for K3.
It makes up far fewer facts than K3 did. On the hard hallucination test with tools, K3 had a 23.4 rate. K3.1 Pro is down to 8.3. That beats Opus 5.5 at 15.0, though GPT-6 Astra is still lower at 5.5. Moonshot is clear that this is a stress test with deliberately hard prompts, so these are not everyday error rates.
3D is close to the top. Without tools on BenchCAD, it sits at 72.8 against Opus 5.5's 73.0. With tools, Opus and Astra pull ahead.
Moonshot also claims a state-of-the-art 64.4 on OfficeQA Pro v2, and big jumps on its internal finance tests. Business operation went from 8.9 on K3 to 37.
Benchmarks only tell you so much. That is why I keep my own Goldie Bench for the jobs I actually run.
📺 Watch: Gemini 3.1 Pro + Kimi K2.5 is INSANE! 🤯
The demos: walking bodies and 3D builds
The most interesting demo in Moonshot's blog is a human walking simulation. It is built on the open-source human-atlas model, with 296 skeletal and 402 muscle components.
At the start, the simulated human collapsed within seconds. After K3.1 Pro worked on it, it walked continuously for 11.2 seconds. Along the way, the model found and fixed 36 shoulder-muscle moment-arm errors.
That is the kind of long, fiddly, multi-step work agents need to be good at.
On the 3D side, Moonshot says K3.1 Pro works with Unity, Unreal Engine and Godot, and builds assets for games, brand films and interactive Three.js scenes. It also powers 3D Make on kimi.ai, where you can preview and edit a model, download it for 3D printing, or export it as a video.
In my video I ran the same build prompts through Kimi and Claude side by side. The gap is small enough that price becomes the deciding factor for a lot of jobs.
📺 Watch: DeepSeek V4 Pro vs Claude Fable 5 vs Grok 4.6
Kimi K3.1 Pro pricing
Here is Moonshot's official API pricing, per million tokens:
- kimi-k3.1-pro: $2.00 input, $10.00 output, $0.20 cached input.
- kimi-k3.1-pro-highspeed: $4.00 input, $20.00 output, $0.20 cached input.
- kimi-k3 (for comparison): $3.00 input, $15.00 output, $0.30 cached input.
So K3.1 Pro is actually cheaper than K3 while scoring higher. Against Opus 5.5, it works out at about half the price for near-identical coding scores.
Caching is automatic. Repeated prompt prefixes are cached for 5 minutes by default, or 1 hour if you ask for it, and a cache write costs $2.00 or $4.00 per million tokens. Pricing is flat, with no extra tier for long context.
One catch: K3.1 Pro unlocks once you have topped up at least $1 on your Kimi API account, and your rate limits grow with your total top-ups.
Not sure where a cheaper model like this fits in your business? Grab a free AI strategy session and we will map it out with you.
When do the open weights land?
Moonshot says the full K3.1 Pro weights will be released by November 4, 2026. As of today, the moonshotai page on Hugging Face still shows Kimi-K3, listed at 2.8T parameters, with no K3.1 Pro repo yet.
Can you run it yourself? Technically yes, once the weights are out. Practically, a 3T-class model needs serious data-centre hardware. For most people, the API or a hosted provider is the way in.
The big win of open weights is choice. Other providers can host it, companies can run it privately, and prices tend to fall as hosts compete.
Where it's weaker
It is not better than Opus 5.5 at everything. Going by Moonshot's own charts:
- Terminal-Bench: still slightly behind Opus 5.5.
- 3D with tools: 92.7 against 96.2 for Opus 5.5 and 95.9 for GPT-6 Astra.
- Hallucination: GPT-6 Astra is still lower in both the tools and no-tools tests.
- Cybersecurity: Moonshot deliberately removed cyber data from training for the open-weight release. On its internal xdev-bench-hard, K3.1 Pro scores 81.0, and Moonshot says its cyber skills trail frontier closed models.
- Web search tool: the API docs say the built-in web search is being updated and is not recommended for production for now.
Speed is the other one. With thinking always on and max effort as the default, answers can take a while. The highspeed version or low effort fixes that, at a cost.
How to use Kimi K3.1 Pro
You have four ways in, per Moonshot:
- Kimi app: on kimi.ai and the iOS, Android and HarmonyOS apps.
- Kimi Work: the desktop app, version 3.1.0 or later, on Windows and Apple silicon Macs.
- Kimi Code: pick the model with the /model command.
- Kimi API: use the model ID kimi-k3.1-pro or kimi-k3.1-pro-highspeed.
The API is OpenAI-compatible. The base URL is https://api.moonshot.ai/v1, and you keep your key in an environment variable called MOONSHOT_API_KEY. A few rules from the quickstart:
- Leave temperature and top_p out. They are fixed at 1.0 and 0.95.
- Set reasoning_effort to low, high or max. Use low when speed matters.
- In tool-call flows, send back the full assistant message, not just the text.
- Images go in as base64, not public URLs.
Zero Data Retention is available for API customers, which matters if you are sending client work through it.
My favourite way to use it is inside an agent. I run Hermes daily with a profile per model, and I walk through that exact setup in Kimi K3.1 Pro Hermes. If you set up the earlier model, my Kimi K3 Hermes Agent guide covers that, and Kimi K3 vs Fable 5 puts the two head to head.
For the bigger picture of how models plug into one system, read my Agent OS guide. Want help picking the right model for each job? Book a free AI strategy session.
FAQ
Is Kimi K3.1 Pro better than Claude Opus 5.5?
Not quite. On Moonshot's numbers it is less than a point behind on Terminal-Bench 4.0 and level on 3D without tools, while making up fewer facts. Opus still leads on 3D with tools. The big difference is price: Kimi is about half.
How much does Kimi K3.1 Pro cost?
$2.00 per million input tokens and $10.00 per million output tokens on Moonshot's API, with cached input at $0.20. The highspeed version is $4.00 input and $20.00 output.
Is Kimi K3.1 Pro open source?
Moonshot says the full weights will be released by November 4, 2026. Until then, you use it through the Kimi apps and API.
What is the context window?
1M tokens, for both the standard and highspeed versions.
Want to plug Kimi K3.1 Pro into real agent workflows and save hours every week? Join the AI Profit Boardroom, or start with a free AI strategy session.
