Kimi K3.1 Pro nearly ties Claude Opus 5.5 at coding, and the weights are coming

For most of the past year, the story of open AI models has been simple: they are good, they are cheap, and they are clearly a step behind the best closed systems from Anthropic, OpenAI and Google. A release from Moonshot AI on 23 October 2026 puts real pressure on that story. The Beijing based lab launched Kimi K3.1 Pro, the first model of its new K3.1 series, and the early numbers place it very close to the top of the field.

On Artificial Analysis, the independent benchmarking site that many developers use to compare models, Kimi K3.1 Pro scores 55 points on the Intelligence Index. That is fourth place out of more than 200 models, behind only Claude Opus 5.5 with 58 points, and Claude Sonnet 5.5 and Muse Wonder Full with 56 each. Its predecessor, Kimi K3, sat at about 44 points. An 11 point jump between two versions of the same base model is unusual at this level of the leaderboard.

There is one important catch, and it is in the headline of the launch itself. Moonshot calls K3.1 Pro its "most capable 3T class open weight model to date", but the weights are not public yet. The company says the full model weights will be released by 4 November 2026. Until then, K3.1 Pro is something you use through Kimi's own apps and API, and Artificial Analysis currently lists it as a proprietary model. If the weights do appear on time, it becomes the strongest open weight model anyone can download.

What Moonshot actually released

Two models launched together. Kimi K3.1 Pro is the main one. Kimi K3.1 Pro Highspeed is a faster variant that Moonshot says runs two to three times quicker, for people who care more about response time than price. Both have a context window of one million tokens, which is enough to hold a large codebase or several long reports at once, and both accept text and images as input.

They are available on Kimi.ai and the Kimi mobile apps, in the desktop app Kimi Work, in the coding tool Kimi Code, and through the Kimi API under the names kimi-k3.1-pro and kimi-k3.1-pro-highspeed. For business customers, administrators have to switch the new model on themselves, because access is off by default at launch. Moonshot also says that zero data retention is available for API customers.

Under the hood, very little changed in size. According to The Decoder, K3.1 Pro keeps the same dimensions as Kimi K3: about 2.8 trillion parameters in total, of which roughly 104 billion are active for any given token. This is a mixture of experts design, where only a slice of the network works on each step, which keeps it practical to run despite the huge total size.

What did change is the training after the base model was built. Moonshot describes K3.1 Pro as a post trained version of K3 with "substantially larger post training compute". In plain terms, the team spent far more effort teaching the existing model through reinforcement learning, across what it calls hundreds of downstream domains, from coding and 3D design to finance. To make that work at scale, Moonshot says it rebuilt much of its training stack, including the reinforcement learning framework, long context inference, sandbox infrastructure and quality checks for the training environments.

Coding: one point behind Claude Opus 5.5

The number that got the most attention on Hacker News and on X is the Terminal-Bench 4.0 score. This benchmark tests coding agents on realistic terminal tasks with limited resources and strict, task specific tests. It is close to what developers actually do when they let an AI agent loose in a project: install things, run scripts, fix what breaks.

Moonshot reports 65.7 percent for K3.1 Pro. Claude Opus 5.5 scores 66.4 percent, GPT-6 Astra 58.2 percent and Claude Fable 5.1 57.9 percent. Other strong open models are far behind: DeepSeek V4.1 Flash at 42.3 percent and GLM-5.3 at 41.8 percent. Kimi K3, the predecessor, managed only about 13 percent on the same test.

On other coding tests the picture is similar. On SWE Marathon v1.1, a long running software engineering benchmark, Moonshot reports 51.7 for K3.1 Pro against 46.3 for Claude Opus 5.5. On DeepSWE v1.1 it reports 75.5, slightly above Opus 5.5 at 74.2, though DeepSeek V4.1 Flash leads that one with 77.0. On Moonshot's own long horizon coding test, Kimi Code Bench 3.0, K3.1 Pro scores 61.2 percent against 55.95 percent for K3.

Moonshot also gives a striking example of long running agent work. In an internal test, K3.1 Pro built a simulator of human walking on top of an open source anatomical model with 296 skeletal parts and 402 muscle parts. Over 167 missions and 420 independent reviews, it analyzed results, found problems and refined the controller, until the simulated human went from collapsing within seconds to walking with alternating strides for 11.2 seconds. When adjusting the reward did not fix the arm movement, the model probed the simulator itself and found 36 errors in how the shoulder muscles were calculated. That is a vendor anecdote, not a benchmark, but it shows the kind of work Moonshot is aiming at.

Agents, office work and 3D

The biggest gains, according to The Decoder's reading of the Artificial Analysis data, are in agentic tasks, where the model has to plan and act over many steps. On GDPval-AA, which measures real world knowledge work tasks with an Elo rating, K3.1 Pro reaches 1,846, a full 312 points above K3. Only Claude Opus 5.5 at 1,866 and Muse Wonder Full at 1,857 are higher. GPT-6 Astra sits at 1,573 on that test.

On OfficeQA Pro v2, which tests analysis of financial data across real documents, Moonshot reports 64.4, the best result of all models in its table, ahead of Claude Opus 5.5 at 62.2. On AutomationBench-AA, a test of agentic workflows in business software, it reports 73.3 against 69.5 for Opus 5.5. On BrowseComp, a hard web research benchmark, it reports 92.1.

3D work is a particular focus. On BenchCAD, which asks a model to rebuild a 3D object from several rendered views by writing CAD code, K3.1 Pro scores 92.7 percent with tools, close to Claude Opus 5.5 at 96.2 percent and GPT-6 Astra at 95.9 percent. Moonshot says the model can write and run scripts for Blender and CAD software directly, and it launched a new feature called 3D Make on Kimi.ai that lets users preview, edit and export 3D models, for example for 3D printing.

There are also areas where K3.1 Pro is clearly not the leader. On GPQA Diamond, a hard science question test, it scores 94.2 against 96.1 for GPT-6 Astra. On ExploitBench, a cyber offense benchmark, it scores 74.7, far below Claude Opus 5.5 at 91.0 and GPT-6 Astra at 100. As we will see, that last gap is partly deliberate.

Cheaper than the model it replaces

Price is where K3.1 Pro becomes hard to ignore for developers. Moonshot cut token prices by about a third compared with K3. Input now costs 2 US dollars per million tokens, down from 3. Output costs 10 dollars per million tokens, down from 15. Cached input costs 20 cents per million tokens, down from 30 cents. The faster Highspeed variant costs twice as much: 4 dollars for input and 20 dollars for output.

There is a hidden cost, though. K3.1 Pro thinks a lot. When Artificial Analysis ran its Intelligence Index, the model generated about 150 million output tokens, compared with a median of about 87 million for models in its class. Artificial Analysis calls it "very verbose". Because it writes so much, the cheaper per token price does not translate into a much cheaper job. Artificial Analysis puts the cost per Intelligence Index task at about 2.11 dollars on its live page, and The Decoder reported a cost per task almost identical to K3. It is also not fast: Artificial Analysis measures about 73.5 output tokens per second through Kimi's API, below the average of comparable models.

Even so, a model that scores within three points of the leader, at a list price of 2 and 10 dollars per million tokens, gives buyers real leverage when they negotiate with closed model providers.

The cyber section: holding back on purpose

The most interesting part of the release is not a benchmark win. It is a long section in Moonshot's technical blog about cybersecurity, and it is unusually candid for a launch post.

Moonshot says that as it scaled up post training, the model's ability to find and exploit software vulnerabilities kept rising. When the team traced where that came from, it found cyber related content inside the general training data and cyber related tasks inside the reinforcement learning mix. Nobody set out to train a hacking model; the skill came along with broad training. In early experiments, Moonshot says, cyber capability would have grown substantially if it had done nothing.

So for the open weight release, Moonshot removed cyber related data from both the mid training stage and the reinforcement learning stage, while keeping the general coding tasks the model needs to be useful. It then measured the effect on an internal test called xdev-bench-hard. Most tasks reproduce recently disclosed vulnerabilities in software like the Linux kernel, Node.js and JavaScriptCore, with mitigations switched on. The agent gets the vulnerable code, a short hint and the binary, and has to achieve code execution and fetch a secret string.

A version of the model trained without the cleanup solved 90.5 percent of those tasks. The released K3.1 Pro solves 81.0 percent. That is a real reduction, but it is still a big step up from Kimi K3 at 33.3 percent. According to The Decoder, DeepSeek V4.1 Flash reaches 85.7 percent on the same test. Moonshot's own conclusion is that general coding and debugging skill carries over into cyber work, even after targeted data removal.

Moonshot lays out two principles for open weights. First, it does not want to release open weights that would push the absolute frontier of offensive cyber capability, and it says K3.1 Pro stays behind closed frontier models that have been public for at least three months. Second, it intends to raise the cyber capability available in open weights step by step, arguing that open models help defenders audit their own code, including old systems that rarely get a proper security review.

This matters because cybersecurity is the topic of the month. Anthropic and others have warned that even mid sized open models can write working exploits, and many teams are scrambling to patch old code that AI tools now find holes in. A Chinese lab publishing its own measurements, including the number it would have reached without restraint, gives outsiders something concrete to check once the weights are out. It also raises a fair question: once the weights are public, anyone can fine tune the model on exactly the data Moonshot removed.

Fewer hallucinations, with a big footnote

Moonshot also claims a large drop in made up facts. On its internal Kimi Hallu Bench Hard, the claim level hallucination rate fell from 23.4 percent for K3 to 8.3 percent for K3.1 Pro. GPT-6 Astra still does better on the same test with 5.5 percent, according to The Decoder.

Moonshot itself adds the footnote: the test is a stress test made of prompts deliberately built to trip models up, so the numbers are not everyday error rates. That is honest, and it is also a reminder of how much of this launch rests on benchmarks the vendor designed and ran itself.

How much to trust the numbers

It is worth being clear about what is independent and what is not. The Intelligence Index score of 55, the fourth place ranking, the speed and the verbosity come from Artificial Analysis, which runs its own tests. Those are the strongest data points here. Most of the other figures, including Terminal-Bench 4.0, SWE Marathon, BenchCAD, the finance results, the hallucination rate and the cyber tests, come from Moonshot's own blog. Moonshot ran its model in its best setting, with reasoning effort set to maximum, and in several cases took competitor scores from their own reports or from public leaderboards. That is normal practice, but it is not the same as a neutral head to head test.

Moonshot also notes some quirks in its footnotes. In a few agent benchmarks, it says it caught other models looking up answers online instead of doing the required analysis, and gave those runs a score of zero. That might be fair, but it shows how much results can depend on the rules each lab applies.

The real test comes with the weights. If Moonshot publishes them by 4 November, independent researchers can run the model on their own hardware, check the cyber claims and see whether it holds up outside Kimi's own harnesses.

Why this release matters

Three things stand out. First, the gap between open and closed models is shrinking again. The Decoder calculates that if K3.1 Pro becomes downloadable as planned, the distance between the best open weight model and the best closed one on the Artificial Analysis index would fall from about 12 points to about 3. Second, the progress came from training, not size. Moonshot kept the same 2.8 trillion parameter model and squeezed out much more capability through reinforcement learning, which suggests other labs can do the same with what they already have. Third, the cyber section sets a precedent. An open weight developer measured how dangerous its own model could have been and published the number next to the one it shipped.

For developers, the practical takeaway is simple: K3.1 Pro is worth testing now through the API, especially for long coding and office automation jobs, with an eye on how many tokens it burns. For everyone else, the date to watch is 4 November. That is when we find out whether a near frontier model really ends up in everyone's hands.

Sources

• Moonshot AI: Introducing Kimi K3.1 Pro (23 October 2026)
• Artificial Analysis: Kimi K3.1 Pro model page
• The Decoder: Chinese AI model Kimi K3.1 Pro is closing in on Anthropic's Claude Opus 5.5 (23 October 2026)