Moonshot AI has released Kimi K3.1 Pro, a post-trained upgrade to its open-weight Kimi K3 that the Chinese lab calls its most capable 3-trillion-parameter-class model yet. The company says the model scores 65.7 percent on Terminal-Bench 4.0, ahead of OpenAI’s GPT-6 Astra at 58.2 percent and just 0.7 points behind Anthropic’s Claude Opus 5.5 at 66.4 percent. That near tie on a demanding test for coding agents is the clearest signal that open-weight models are beginning to close the capability gap with leading proprietary systems.
The release arrives as Kimi K3.1 Pro, along with a faster companion variant, becomes available through Kimi.ai, Kimi Work, Kimi Code and the Kimi API. It retains the same underlying size as K3, with 2.8 trillion total parameters, 104 billion active parameters and a context window of one million tokens. It also accepts both text and images. The real change sits in post-training, where Moonshot says it used substantially more compute and built joint reinforcement learning across hundreds of downstream domains.
Kimi K3.1 Pro Performance on Agentic Coding
Terminal-Bench 4.0 is the benchmark drawing the most attention because it evaluates coding agents on realistic terminal tasks where speed, resource limits and reliable tool use all matter. Kimi K3.1 Pro’s 65.7 percent result places it behind Claude Opus 5.5 but well ahead of GPT-6 Astra. For comparison, the earlier Kimi K3 managed only 13 percent on the same test, indicating a dramatic improvement in multi-step execution rather than a modest refinement.
Moonshot also reports 61.2 percent on Kimi Code Bench 3.0 Long-horizon, its internal test covering extended execution, tool use and end-to-end problem solving, compared with 55.95 percent for K3. Third-party evaluation from Artificial Analysis placed Kimi K3.1 Pro fourth on its Intelligence Index with 55 points, an 11-point gain over K3. Only Claude Opus 5.5 at 58, Claude Sonnet 5.5 at 56 and Muse Wonder Full at 56 ranked higher. On the GDPval-AA benchmark, the model reached 1,846 Elo, trailing Claude Opus 5.5 at 1,866 and Muse Wonder Full at 1,857.
If Moonshot publishes the full weights as planned by November 4, 2026, Kimi K3.1 Pro would become the highest-scoring open-weights model available. Based on current index figures, that would narrow the gap between open and proprietary models from 12 points to 3. Even with unchanged competitors, this would mark a meaningful redistribution of capability from closed labs toward teams that can inspect, host and adapt the model themselves.
Lower Token Prices and a Faster Variant
Moonshot reduced API prices by 33 percent for the standard model. Input now costs $2 per million tokens, down from $3, while output is $10, down from $15. Cache hits are priced at $0.20 per million tokens, down from $0.30. Despite the reduction, the cost of one Intelligence Index task is about $1.98, nearly identical to K3 at $2.00, because Kimi K3.1 Pro generates roughly 14 percent more output tokens.
A second option, Kimi K3.1 Pro Highspeed, runs two to three times faster with equivalent capabilities. According to one report, it is priced at roughly twice the standard rate for most usage, although cache-hit input pricing is similar. Both versions offer low, medium and max reasoning settings, giving developers a direct trade-off between latency, depth and cost. Enterprise administrators can enable the model for their organizations, with access disabled by default at launch.
Moonshot has not published the full training bill. It says the company optimized the reinforcement learning framework, long-context inference, training efficiency, sandbox infrastructure, rollout scheduling and a shared platform for environment quality checks. The emphasis on infrastructure suggests Moonshot views the release as an engineering milestone rather than simply a larger training run. Without a public FLOP count, dataset size or training duration, outside researchers will have limited ability to verify exactly how the gains were produced.
Reduced Hallucinations and Stronger Office Performance
Moonshot reports that claim-level hallucinations on its internal Kimi Hallu Bench Hard fell from 23.4 percent with K3 to 8.3 percent. The proportion of responses containing at least one factual error also dropped to 42.4 percent of the predecessor’s rate. GPT-6 Astra still records 5.5 percent on the same benchmark. Moonshot stresses that this is a deliberately difficult stress test and not a measure of everyday error rates, a caveat worth keeping in view when interpreting the improvement.
Office and finance benchmarks show the largest gains from the new post-training work. The model recorded 64.4 on OfficeQA Pro v2, which Moonshot describes as a state-of-the-art result for that evaluation of complex fiscal data analysis across real financial documents. Internal finance results include a 7.2-point gain on Financial Modeling and an 11.6-point gain on Financial Research. Business Operation performance rose from 8.9 to 37, roughly four times the predecessor’s score. These results come from Moonshot’s own testing, so independent validation will be important.
3D Modeling, CAD and Game Engine Work
Kimi K3.1 Pro also expands beyond conventional text reasoning into design and spatial tasks. On BenchCAD, which tests 3D reconstruction from multiple views and CAD code generation, the model scored 92.7 percent with tools, close to Claude Opus 5.5 at 96.2 percent and GPT-6 Astra at 95.9 percent. Without tools, it reached 72.8 percent, almost exactly matching Claude Opus 5.5 at 73.0 percent. That small difference in tool-free performance is notable given the size and open-weight status of the model.
Moonshot says the system can work directly with Unity, Unreal Engine and Godot, handling both narrow fixes and complete project builds. It is also positioned for parametric modeling of mechanical parts, architectural structures and configurable products. The model can write and execute scripts for Blender and CAD software without relying on MCP connectors, a practical advantage for workflows where connector setup would otherwise add friction.
A new feature called 3D Make on Kimi.com and the Kimi app lets users preview, edit and export 3D models online, including files prepared for 3D printing. Artificial Analysis reports that Kimi K3.1 Pro sits at the top of its Pareto frontier, meaning no comparable model in its size class produces better results under the same efficiency assumptions. For teams evaluating whether a proprietary frontier model is still necessary, that combination of capability, context length and price is becoming harder to ignore.
Cybersecurity Capabilities Receive Special Scrutiny
Moonshot devoted a full section of its technical blog post to cybersecurity, an area where stronger models create obvious dual-use concerns. The company observed that vulnerability research ability kept improving as post-training compute increased, which it attributed to cyber material embedded in the general mid-training corpus and related tasks included in the reinforcement learning environment mix. That finding illustrates how capabilities can emerge unintentionally when broad training data and broad task environments are combined.
For the open-weight release, Moonshot removed cyber-related data from both training phases. On its internal xdev-bench-hard suite, which reproduces recently disclosed flaws in software including the Linux kernel, Node.js and JavaScriptCore, the uncleaned model solved 90.5 percent of tasks. The cleaned Kimi K3.1 Pro solved 81.0 percent. DeepSeek-V4.1-Flash reached 85.7 percent, while Kimi K3 scored 33.3 percent. The cleaning did not eliminate the capability, but it reduced measured performance by 9.5 percentage points.
This approach raises questions that benchmark tables cannot answer. How much of the model’s security knowledge remains after filtering, and what techniques might restore it? Moonshot’s decision to document the change at least provides a window into the trade-off between model utility and responsible release practices. Other open-weight labs now face a similar choice between withholding models that can find real vulnerabilities and shipping them with security-related content removed.
What Kimi K3.1 Pro Means for the Open-Weight Race
Kimi K3.1 Pro is not a universal replacement for Claude Opus 5.5. It still trails on Terminal-Bench 4.0, the Artificial Analysis Intelligence Index and hallucination testing, and several of its strongest office and finance results come from internal evaluations. Long-context performance under sustained real-world load also needs independent testing rather than reliance on window size alone.
Even so, the direction is difficult to dismiss. The model combines near-frontier agentic coding results, a one-million-token context window, lower API prices and planned open weights at a scale where few comparable systems are available. Moonshot says Kimi K3.1 Pro is its strongest 3-trillion-class open-weight model yet, and if the promised release arrives on schedule, the gap separating open systems from proprietary leaders will be measured in single-digit points on some indexes rather than double digits. That would make Kimi K3.1 Pro a defining release for organizations weighing whether they need closed models at all.