Kimi K3.1 Pro: Open-Weight Coding Models Are Closing the Gap

Moonshot dropped this one at 12:15 PM Central on Friday: "Introducing Kimi K3.1 Pro: our most capable 3T-class open-weight model." I read the announcement, then the tech blog, then the API docs, and I'm honestly a little giddy about it. Not because everyone should switch tomorrow, but because of what the chart says about open models.

Open-weight models used to be what you ran when you couldn't afford the good one. If Moonshot's numbers hold up, that line is getting blurry.

Kimi K3.1 Pro is a post-trained version of Kimi K3, Moonshot's open 3T-class model from this summer, trained with a lot more reinforcement learning on top. According to the launch post and the API quickstart:

• Size and context: 2.8T total parameters, 104B active per token, multimodal (text, images and video), 1M-token context.
• Two versions: kimi-k3.1-pro and kimi-k3.1-pro-highspeed, which Moonshot says runs 2-3x faster.
• Where it runs today: Kimi.ai and the mobile apps, the Kimi Work desktop app, Kimi Code (pick it with /model) and the Kimi API.
• Always thinking: reasoning can't be turned off. You set reasoning_effort to low, high or max, and max is the default.

Pricing on the Kimi API Platform is $2 per million input tokens and $10 per million output, with cache hits at $0.20. Highspeed costs $4 and $20. For comparison, K3 is listed at $3 and $15, so the new model is cheaper than the one it replaces. You need a minimum $1 top-up to unlock it.

Terminal-Bench 4.0 tests whether an agent can actually get work done in a terminal: software engineering, system administration, ML and security tasks, graded by strict tests. That's the kind of work I hand to agents in Cursor all day, so it's the number I care about most.

In Moonshot's chart, K3.1 Pro scores 65.7%, against 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra. The previous open-weight leaders in the same chart sit way back: DeepSeek V4.1 Flash at 42.3% and GLM-5.3 at 41.8%. That's a jump of more than 20 points for open models in one release.

Here's the fine print, and it matters. These are vendor-reported numbers. Moonshot ran K3.1 Pro itself, inside the Claude Code harness at max reasoning effort. Its footnotes say competitor scores are the higher of each lab's own published figure or the official Terminal-Bench leaderboard. On that leaderboard right now, Opus 5.5 in Claude Code shows 64.85% and GPT-6 Astra in Codex shows 58.18%. K3.1 Pro isn't on it yet.

Artificial Analysis rolled out version 2.0 of its Coding Agent Index less than a day before this launch. It's an equal-weight average of three benchmarks, FrontierCode v1.1, Terminal-Bench 4.0 and SWE-Atlas-QnA-AA, run by Artificial Analysis itself across 290 tasks (methodology). It scores a model and its agent harness together, which is how we actually use them.

• Top of the board: Claude Code with Opus 5.5 at 61.7.
• Best open-weight entry: GLM-5.3 in Opencode at 48.9.
• Kimi K3 in Kimi Code CLI: 41.4, with 21.2% on the Terminal-Bench 4.0 component.

K3.1 Pro hasn't been added yet. That's the comparison I'm waiting for. Artificial Analysis runs its own setup, so its scores won't match Moonshot's directly, but if K3.1 Pro lands anywhere near the top of that board, the "open models are a generation behind" story is done.

The tech blog has the fun demos. Moonshot says K3.1 Pro can build full projects in Unity, Unreal Engine and Godot, and shows examples it says the model made end to end, including a celestial palace in Unreal Engine 5.8 and an action-adventure game in Godot 4.7.2 with five regions, combat and a boss fight.

Those are vendor demos, picked by the vendor. Still, I'm here for it. I build WordPress sites and Next.js apps, not games, but a model that can hold a whole Godot project in its head is a model that can probably hold a messy WooCommerce install in its head too.

Moonshot says it removed cyber-related training data before the open release. On its in-house exploit benchmark, a checkpoint trained without those cuts solved 90.5% of tasks, and the released model solves 81.0%. Moonshot's stated rule is not to release open weights that push the frontier of offensive capability. After a month of writing about AI-found vulnerabilities, I'm glad someone is publishing numbers on this.

I use Cursor every day, and my current AI coding setup leans on the big closed models for the hard stuff. Open models have mostly been my side project, like the local coding model I tried and the MiniMax M3.1 Flash open weights I wrote about today.

A frontier-class open model changes the math for client work. You choose where it runs and lock the version. I won't run a 2.8T model on my desk, but once the weights are out, other providers can host it and compete on price.

What I'll do first is try it on the Kimi API with some real PHP and WordPress tasks, then see where it lands when independent results come in. Benchmarks are a starting point. My own repos are the real test.

The weights aren't public yet, so nobody outside Moonshot can test the open release. There's no Hugging Face repo for K3.1 Pro as of this writing, and Moonshot hasn't named the license. K3 shipped under the Kimi K3 License, which is permissive but asks very large commercial users to sign a separate agreement or show "Kimi K3" in their product. I'd expect something similar, but I'm not going to guess. I'll update this post when the weights and license land.

Open models went from "nice for the price" to "within a point of the best closed model," at least by the vendor's own scoreboard. Even with the asterisks, that's a wild thing to watch happen. November 4 is the date to watch.

If you're trying to figure out which AI tools make sense for your site or app, let's talk.

Is Kimi K3.1 Pro open source?

Not yet. Moonshot calls it open-weight and says the full weights will be released by November 4, 2026. It hasn't named the license, and open weights aren't the same as open source.

How does it compare to Claude Opus 5.5 and GPT-6 Astra?

On Terminal-Bench 4.0, Moonshot reports 65.7% for K3.1 Pro, 66.4% for Opus 5.5 and 58.2% for GPT-6 Astra. Those are Moonshot's numbers. Independent results from Terminal-Bench and Artificial Analysis aren't out yet.

How much does the API cost?

$2 per million input tokens and $10 per million output for K3.1 Pro, and $4 and $20 for Highspeed. Cached input is $0.20 per million for both.

How big is the context window?

1 million tokens, with flat pricing across the whole window.

Can I use it in Cursor?

Moonshot hasn't announced Cursor support. Today you can use it in Kimi Code or through the Kimi API.

• Kimi (Moonshot AI) on X, "Introducing Kimi K3.1 Pro" (Oct 23, 2026), with benchmark chart
• Kimi blog, "Introducing Kimi K3.1 Pro" tech blog and benchmark footnotes
• Kimi API Platform, Kimi K3.1 Pro quickstart
• Kimi API Platform, model pricing
• Terminal-Bench 4.0 leaderboard
• Artificial Analysis, Coding Agent Index v2.0 leaderboard
• Artificial Analysis, Coding Agent Index v2.0 methodology
• Hugging Face, moonshotai/Kimi-K3 model card
• Hugging Face, Kimi K3 License
• Hugging Face, zai-org/GLM-5.3 model card