Moonshot Unveils Kimi K3.1 Pro, Claiming a Coding Edge Over GPT-6

Date:

Moonshot AI has launched Kimi K3.1 Pro, a new model the Chinese startup calls its most capable open-weight model to date. Moonshot says the model beats OpenAI’s GPT-6 Astra on a major coding-agent test, scoring 65.7 percent against 58.2 percent on Terminal-Bench 4.0, though it still trails Anthropic’s Claude Opus 5.5 at 66.4 percent on the same test.

Those numbers are Moonshot’s own, not independent results, and that caveat runs through this whole story. K3.1 Pro launched on October 23, 2026, and Moonshot says its full model weights will be released by November 4. Until that release happens and outside researchers can test the model themselves, every benchmark claim below should be read as a vendor claim, not a verified fact.

What Kimi K3.1 Pro Actually Is

The first thing to understand is that K3.1 Pro is not a brand-new base model. Moonshot describes it as a post-trained version of Kimi K3, which launched on July 16, 2026. The base model did the heavy lifting in July. This release is mostly a post-training story, built with what Moonshot calls substantially larger post-training compute and reinforcement learning across hundreds of domains.

In plain terms, Moonshot took the existing K3 model and trained it much harder on real tasks: coding, tool use, document analysis, and multi-step problem solving. The company says it optimized every layer of that training stack, from the reinforcement learning framework to long-context inference, training efficiency, sandbox infrastructure, rollout scheduling, and a shared platform for checking environment quality.

The headline specs, as listed in Moonshot’s launch post, are a 2.8-trillion-parameter total with 104 billion active parameters, in a multimodal mixture-of-experts design, with a 1-million-token context window. Mixture-of-experts means only a fraction of the model activates for any given token, which is how a model can be huge on paper while staying practical to run. The 1-million-token window means it can, in principle, hold very large codebases or document sets in a single session.

A developer workspace with code on large monitors and an AI assistant interface, illustrating AI-assisted software development

Where Moonshot Claims a Coding Edge Over GPT-6

The coding claim is the headline, so it deserves a careful look. Terminal-Bench 4.0 grades coding agents on realistic terminal tasks with strict per-task tests. Moonshot reports 65.7 percent for K3.1 Pro, 66.4 percent for Claude Opus 5.5, and 58.2 percent for GPT-6 Astra. On that test, K3.1 Pro sits just behind Opus and well ahead of GPT-6 Astra, a gap of 7.5 points.

Moonshot also reports a lead on SWE Marathon, at 51.7 percent against 46.3 percent for Opus 5.5 and 42.5 percent for GPT-6 Astra. On Agents’ Last Exam, another agent test, Moonshot reports 43.2 percent for K3.1 Pro against 42.6 percent for Opus and 32.8 percent for GPT-6 Astra. On its own internal Kimi Code Bench 3.0 Long-horizon test, K3.1 Pro scored 61.2 percent, up from 55.95 percent for K3, though an internal test designed by the vendor is the weakest form of evidence here.

The honest summary is narrower than the headline. Moonshot claims an edge over GPT-6 Astra on several coding and agent tests, but not over Opus 5.5 overall. On Terminal-Bench 4.0, the most cited test, Opus still leads by 0.7 points. Anyone choosing a coding model on these numbers alone would pick Opus first, K3.1 Pro second, and GPT-6 Astra third, and even that ranking rests on Moonshot’s own chart.

There is also a method footnote that matters. Moonshot’s footnotes say K3.1 Pro ran inside the Claude Code harness for most coding tests, while many rival scores were taken from public leaderboards or the rivals’ own reports. Different harnesses can shift scores, so a head-to-head comparison run under identical conditions could look different. Independent testing after the weights release will be the real check.

Where Kimi K3.1 Pro Loses Ground

A balanced launch story has to include the losses, and K3.1 Pro has clear ones. The biggest is science. On Terminal-Bench Science 0.1, Moonshot reports 38.1 percent for K3.1 Pro, against 68.1 percent for GPT-6 Astra and 63.3 percent for Opus 5.5. That is not a close gap. It suggests the model is much stronger at coding and document work than at scientific problem solving, at least on this test.

The picture is mixed on 3D and design work. On BenchCAD, which asks models to rebuild 3D objects as CAD code, Moonshot reports 92.7 percent with tools for K3.1 Pro, behind Opus 5.5 at 96.2 percent and GPT-6 Astra at 95.9 percent. Without tools, K3.1 Pro scores 72.8 percent, close to Opus at 73.0 percent. A 3D Make feature is live on kimi.ai and the Kimi app, letting users preview and edit 3D models online and download them ready for 3D printing, which may matter more to everyday users than the benchmark gap.

On finance, Moonshot claims a state-of-the-art 64.4 on OfficeQA Pro v2, a test of analysis across real financial documents. That claim is again Moonshot’s own, and OfficeQA Pro v2 is not as widely cited as the coding tests, so readers should treat it as a claim to verify, not a settled ranking.

On factual accuracy, Moonshot reports progress on its own hardest test. Claim-level hallucinations on the internal Kimi Hallu Bench Hard fell from 23.4 percent for K3 to 8.3 percent for K3.1 Pro. GPT-6 Astra still posts a lower 5.5 percent on the same test, according to Moonshot’s chart. The direction is good, but the model is not leading on accuracy.

A 3D printer creating a small geometric object in a workshop

Why Moonshot Removed Cyber Data Before the Open Release

One of the most interesting parts of this launch has nothing to do with benchmarks. Moonshot says it removed cyber-related data from mid-training and reinforcement learning for the open-weight release. The reason is safety: the company does not want the downloadable weights to be easily usable for offensive cyber work.

Moonshot quantified the cost of that choice in its own terms. On its internal xdev-bench-hard test, an unfiltered checkpoint solved 90.5 percent of tasks at pass@3, while the shipped K3.1 Pro solves 81.0 percent. Moonshot describes that 9.5-point gap as its own measure of what the removal took away, and says the shipped model stays behind frontier proprietary models on offensive cyber tasks.

This is a trade-off worth understanding. Open-weight releases let anyone download and run a model, which is powerful for research and for builders who want control. It also means the developer cannot control what users do with it after download. Removing cyber data is Moonshot’s attempt to keep the open release useful for coding and office work while limiting its misuse potential. Whether that balance holds will depend on what independent researchers find once the weights are out.

Kimi K3.1 Pro Highspeed, Pricing, and Where to Try It

Alongside K3.1 Pro, Moonshot launched Kimi K3.1 Pro Highspeed, which the company says has the same abilities but runs two to three times faster. Both models are live now in the Kimi app, the Kimi Work desktop app, the Kimi Code coding tool, and the Kimi API. Kimi Enterprise administrators can enable K3.1 Pro for their organizations, with access off by default at launch.

Pricing reports put K3.1 Pro at $2 per million input tokens and $10 per million output tokens, on a flat pay-as-you-go basis with no tiering by context length, according to API coverage of the launch. Cache hits are billed at one-tenth of the cache-miss price for the Kimi K3 family. Those figures come from launch coverage, not from a price list quoted in Moonshot’s own launch post in the sources reviewed, so users should confirm current pricing on the Kimi API page before budgeting a project around it.

The key date is November 4. Moonshot says the full model weights will be released by then. Independent AI commentator Nathan Lambert, who writes the Interconnects newsletter on open models, wrote on X that the benchmark scores “look to be clearly the strongest current open weight model, if the weights are released as they claim by November 4th.” That “if” is doing a lot of work. Open-weight promises have slipped before across the industry, and the release itself, not the announcement, is what researchers can test.

How Does Kimi K3.1 Pro Fit the Wider AI Race?

K3.1 Pro lands in a crowded field where open-weight models are closing the gap on closed ones, at least on coding. That trend matters because open weights change who can build on a model. A startup, a university lab, or an individual developer can download the weights, fine-tune them, and run them on their own hardware, without depending on a vendor’s API or pricing.

The UK’s recent move into AI biosecurity shows the other side of that same coin: governments are now building labs and partnerships around AI safety and capability. For context on how public institutions are responding to frontier AI, see how the UK launched an AI biosecurity lab with OpenAI and Anthropic. The two stories together show an industry splitting into open builders and guarded institutions, both moving fast.

For users, the practical question is simpler: what should you actually use K3.1 Pro for? Based on Moonshot’s own claims, coding assistance, multi-step agent tasks, and financial document analysis are its strengths. Scientific research and tasks where factual accuracy is critical are areas to approach with more caution, given the science gap and the hallucination rate. And for any serious project, the wise move is to wait for independent benchmarks after November 4 before switching tools on the basis of a vendor chart.

A network of connected nodes and code symbols on a dark background, illustrating an open-weight AI model ecosystem

What to Verify Before November 4

Three checks will tell readers whether K3.1 Pro lives up to its launch. First, does the weight release actually arrive by November 4, complete and usable, or does it slip or arrive with restrictions? Second, do independent researchers reproduce the Terminal-Bench and SWE Marathon scores under identical conditions, or do the numbers fall when the vendor harness advantage disappears? Third, does the cyber-data removal hold up in practice, or do researchers find the open model still capable of the offensive tasks Moonshot says it removed?

Until those checks happen, the fair verdict is this: Moonshot has announced a strong open-weight coding model with credible-looking vendor numbers, a clear science weakness, and a thoughtful safety trade-off. That is a story worth following. It is not yet a proven new leader, and any article that calls it one without the vendor caveat is overstating the evidence.

LEAVE A REPLY

Please enter your comment!
Please enter your name here

Share post:

Subscribe

Popular

More like this
Related

The Body on Netflix: Release Date, Cast and Story

The Body premieres on Netflix today, Wednesday, November 11,...

India Cuts Cancer Drug Prices by Up to 70%: 137 Medicines Get Cheaper

Cancer medicines in India got cheaper on November 9,...

GTA 6 Pre-Loading Begins: How to Download and Release Date

Grand Theft Auto VI pre-loading is live. Rockstar opened...

NFL Week 10 Schedule 2026: Dates, Times and TV for All 14 Games

NFL Week 10 runs from Thursday, November 12 to...