News update: this post was researched and drafted automatically from the sources listed at the end, and published without a human edit. If you spot a mistake, write to me and I will fix it.
Moonshot AI has released Kimi K3.1 Pro, a post-trained version of its earlier Kimi K3 model. The new model keeps the same size: 2.8 trillion total parameters, 104 billion active parameters, and a 1 million token context window. Moonshot says full model weights will be released by November 4. A faster variant, Kimi K3.1 Pro Highspeed, is also out and runs 2 to 3 times faster at twice the price, except for input cache hits, which cost the same.
The larger story is that an open weights model is now close to the top proprietary frontier. On Artificial Analysis, K3.1 Pro scores 55 points, which puts it fourth, 3 points behind Claude Opus 5.5 and above GPT-6 Astra and Gemini 4 Argon. For someone running a small delivery team, this matters because the cost and control equations may shift if the weights actually land.
In one line: Moonshot AI's Kimi K3.1 Pro is a 2.8 trillion parameter mixture of experts model with a 1 million token context, scores 65.7 percent on Terminal-Bench 4.0, and promises open weights by November 4, closing the gap with proprietary frontier models to 3 points on Artificial Analysis.
What the numbers say
Key benchmark numbers from Moonshot and Artificial Analysis:
- Terminal-Bench 4.0: 65.7 percent, ahead of GPT-6 Astra at 58.2 percent and just behind Claude Opus 5.5 at 66.4 percent. K3 managed 13 percent.
- GDPval-AA: 1,846 Elo, 312 points above K3. Claude Opus 5.5 is at 1,866.
- Artificial Analysis Intelligence Index: 55 points, up 11 points from K3's 44. Only Claude Opus 5.5, Claude Sonnet 5.5, and Muse Wonder Full rank higher.
- Kimi Hallu Bench Hard, Moonshot's internal hallucination stress test: claim-level errors fell from 23.4 percent for K3 to 8.3 percent. GPT-6 Astra scores 5.5 percent.
These are vendor reported and third party benchmark numbers, not production guarantees. But the jump on Terminal-Bench is large enough to notice.
Coding and agentic work
Moonshot says K3.1 Pro got substantially more post-training compute, with joint reinforcement learning across hundreds of downstream domains. The model handles quick fixes and full project builds. It works with Unity, Unreal Engine, and Godot, and it can write and run scripts for Blender and CAD software without MCP connectors.
One internal test stands out. Moonshot reports that K3.1 Pro built a human walking dynamics simulator on top of an open source anatomical model with 296 skeletal components and 402 muscle components. Across 167 missions and 420 independent reviews, it diagnosed problems and iteratively refined the controller. When reward shaping failed to improve arm motion, the model ran open-loop probes, found and corrected 36 errors in shoulder-muscle moment-arm calculations, and retrained the controller. The simulated human went from collapsing within seconds to walking continuously for 11.2 seconds.
This is the kind of long horizon agentic coding that matters for client work. It is not just code completion. It is running a loop, reading results, and fixing the underlying model.
Prices, weights, and the cybersecurity note
The API price is down 33 percent. Input costs $2 per million tokens, down from $3. Output costs $10 per million, down from $15. Cache hits cost $0.20, down from $0.30. Simon Willison notes the US dollar pricing is the same as GPT-6.1 Sol and half the price of Claude Opus 5.5.
The Highspeed variant is 2 to 3 times faster at twice the price, except for input token cache hits, which stay the same.
Moonshot also spent a section of its announcement on cybersecurity. The company says it stripped cyber-related data from training. On its internal xdev-bench-hard, the uncleaned version solved 90.5 percent of tasks. After cleaning, K3.1 Pro solves 81.0 percent. Moonshot says no open weights model should push the absolute frontier of offensive cyber capability, while still allowing defensive uses like auditing legacy systems.
That is a deliberate tradeoff. It may matter for regulated clients who worry about what open models can do.
What I would try
I would not switch a client pipeline yet. The weights are promised for November 4 but not delivered. The model still trails Claude Opus 5.5 by 3 points on Artificial Analysis and by less than a point on Terminal-Bench. Hallucination on Moonshot's own stress test is still above GPT-6 Astra.
What I would do is run a small internal eval. Take a few long-horizon coding tasks from past client work, the kind that fail when a model loses context or skips a tool call. Run them through the Kimi API and compare against whatever we use now. That is the same long-horizon agentic build I described in A brief goes in, software comes out.
If the weights land and the benchmarks hold, the more interesting path is self-hosting for regulated clients where data cannot leave. At $2 per million input tokens and the possibility of open weights, the math could shift for small teams. For now, just watch.