Kimi K3.1 Pro: Benchmarks, Agentic AI, 3D Tools and Pricing

Kimi K3.1 Pro: Benchmarks, Agentic AI, 3D Tools and Pricing

The identity behind the mysterious AI model “Kibitz,” which attracted attention through anonymous Arena blind tests, has been revealed. Moonshot AI officially introduced Kimi K3.1 Pro on October 23, 2026, positioning it as the first model in its new K3.1 series and a substantial upgrade to the original Kimi K3.

The model’s most notable improvements center on autonomous software engineering, tool usage, 3D creation, financial research, and long-horizon problem-solving. Rather than relying exclusively on a larger pretrained model, Moonshot says it invested heavily in reinforcement learning (RL) and post-training infrastructure to improve the model’s ability to execute complex tasks, diagnose failures, and deliver working results.

Kimi K3.1 Pro combines approximately 2.8 trillion total parameters, 104 billion active parameters, and a native one-million-token context window. A faster variant, Kimi K3.1 Pro Highspeed, is also available, with Moonshot claiming two to three times the output speed of the standard model.

The early benchmark results are competitive with several proprietary frontier models. Artificial Analysis assigned K3.1 Pro an Intelligence Index score of approximately 55, placing it fourth in its October 2026 comparison snapshot. The model also recorded substantial gains in coding-agent and business-workflow evaluations.

However, the distinction between vendor-reported and independently measured results matters. Some of the most impressive demonstrations come from Moonshot’s internal tests, while other results have been evaluated by third parties. As of October 25, the full model weights have not yet been released, although Moonshot has committed to publishing them by November 4, 2026.

🚀 Kimi K3.1 Pro: What Is New?

#

Kimi K3.1 Pro builds on the Kimi K3 foundation introduced in July 2026. The original K3 established Moonshot’s large-scale mixture-of-experts approach, combining a large total parameter count with a smaller active parameter subset for each token.

The K3.1 update focuses on improving what the model can accomplish after receiving an instruction. This includes navigating codebases, using development tools, constructing interactive applications, generating 3D assets, and iteratively correcting errors.

Core Specifications

#

The combination of total and active parameters is important. A 2.8-trillion-parameter MoE model does not activate every parameter for every token. Its expert-routing architecture enables a subset of the network to participate in each computation.

This makes total parameter count an incomplete indicator of inference cost or performance. The actual resource requirements depend on expert routing, model implementation, context length, inference infrastructure, and the amount of work generated by the model.

A Million-Token Context for Long-Horizon Tasks

#

Kimi K3.1 Pro natively supports a one-million-token context window. This provides room for large codebases, lengthy technical documents, substantial financial records, and extended agent execution histories.

For software engineering, long context can help the model retain relationships across files and understand how changes affect a larger system. For financial analysis, it can support retrieval and comparison across multiple source documents. For agentic workloads, it can preserve observations, intermediate results, and debugging history across longer tasks.

However, a large context window does not guarantee that every detail will be retrieved or used correctly. Context quality, retrieval strategy, attention behavior, and task planning remain important factors.

🧠 Reinforcement Learning Is the Core of the K3.1 Upgrade

#

Moonshot describes Kimi K3.1 Pro as a heavily post-trained version of Kimi K3. According to the company’s official technical announcement, the team substantially increased post-training compute and redesigned multiple parts of its reinforcement-learning infrastructure.

The engineering effort covered the RL framework, long-context inference, training efficiency, sandbox infrastructure, rollout scheduling, and environment-quality verification.

The objective was to improve the model’s performance across hundreds of downstream domains while making the returns from additional training compute more consistent.

From Generating Answers to Completing Tasks

#

Traditional language-model evaluations often emphasize whether a model produces a correct answer to a self-contained prompt. Autonomous agents face a different challenge: they must plan a sequence of actions, operate tools, observe intermediate results, and recover from failures.

For example, an agent asked to repair a software project may need to:

1. Inspect the repository and identify the relevant components.
2. Reproduce the reported failure.
3. Determine which code or configuration caused it.
4. Modify the implementation.
5. Run tests and analyze their output.
6. Repeat the process until the failure is resolved.

The final result depends on the entire execution sequence, not simply the quality of the first generated response.

Moonshot’s RL strategy aims to improve these behaviors by training against task outcomes and feedback from execution environments. The company’s reported results suggest that long-horizon coding and tool-use tasks benefited particularly strongly.

For broader context on how reinforcement learning and post-training infrastructure are reshaping coding agents, see KAD’s analysis of Cursor Composer 2.5 and AI cost-performance economics.

📊 Benchmark Results: Kimi K3.1 Pro Approaches Frontier Models

#

Artificial Analysis reported an Intelligence Index score of approximately 55 for Kimi K3.1 Pro, an increase of 11 points over Kimi K3. Its October 2026 comparison placed the model fourth, behind Claude Opus 5.5, Claude Sonnet 5.5, and Muse Wonder Full.

The result is notable because it places K3.1 Pro close to leading proprietary models while Moonshot prepares to release its weights.

Nevertheless, the Intelligence Index combines multiple evaluations. A high aggregate score does not imply that a model leads in every category, and rankings can change as new models and benchmark results are added.

Selected Benchmark Results

#

The table combines vendor-reported and third-party evaluation results. In particular, Moonshot reported 65.7% on Terminal-Bench 4.0, while Artificial Analysis measured 67.2% in its own evaluation. These numbers should not be treated as interchangeable because the evaluation setup and execution conditions can differ.

Moonshot also reported that K3.1 Pro achieved the highest score in 16 of 47 benchmark categories in its published comparison, including three of its own internal benchmarks. The company’s benchmark suite includes coding, scientific reasoning, financial analysis, knowledge work, and agentic workflows.

Agentic Coding Shows the Largest Gains

#

The most striking improvement appears in tasks requiring sustained execution rather than isolated knowledge retrieval.

On GDPval-AA, K3.1 Pro reportedly reached 1,846 Elo, a 312-point increase over K3. This evaluation measures performance on economically valuable professional tasks.

On Terminal-Bench 4.0, the difference from K3’s reported 13% score is especially large. The benchmark evaluates coding agents working through realistic terminal tasks under strict task-specific validation.

AutomationBench-AA, which evaluates SaaS workflow automation, also improved from 58% to 73%.

These gains fit Moonshot’s stated training strategy: reinforcement learning was designed to improve tool use, multi-step execution, and task completion across diverse environments.

By comparison, the increase on Humanity’s Last Exam was more modest, rising from approximately 47% to 52–52.5%. The difference suggests that improvements are not uniform across all benchmark categories.

Where Gaps Remain

#

K3.1 Pro does not dominate every evaluation.

Moonshot’s published results indicate remaining gaps on particularly difficult reasoning and program-reproduction tasks such as HLE-Diamond and ProgramBench. Results in these categories depend on the precise test definitions and evaluation settings, but they illustrate why a single aggregate intelligence score should not be treated as a complete measure of capability.

Long-context performance is another area where scores must be interpreted carefully. A model can accept a million-token input without necessarily outperforming smaller-context models on every task involving extensive information.

The practical question is whether K3.1 Pro can consistently convert its context capacity and reasoning effort into correct, verifiable results.

💻 Coding Agents: From a Single Brief to a Runnable Application

#

Moonshot’s central product claim is that K3.1 Pro can deliver a complete, runnable project from a natural-language brief.

The model is designed to work across code generation, repository modification, web development, tool execution, and iterative debugging. According to Moonshot, it can use tools such as Unity, Unreal Engine, and Godot alongside web search and 3D asset generation to create interactive experiences.

This is broader than producing a code snippet. A completed project may require multiple source files, asset references, configuration settings, dependency management, runtime checks, and a reproducible launch procedure.

Animated SVG: A Five-Layer Character Challenge

#

One hands-on test adapted the familiar pelican-riding-a-bicycle challenge into a more complex animation.

The prompt asked K3.1 Pro to produce an animated SVG showing a game controller riding a Cthulhu-like creature, which rides an elephant, which rides a pelican, which rides a bicycle.

Each character needed a distinct movement pattern. The elephant had to walk on four legs, the tentacled creature had to crawl, and the pelican had to pedal the bicycle while the stacked hierarchy remained visually coherent.

The reported result was a 41KB SVG containing the requested animated elements. The model also supplied an unsolicited backstory titled “Five Layers of Trust.”

This is a useful qualitative test of structured generation because it requires multiple related visual elements to coexist in one output. However, one successful example does not establish general SVG reliability or guarantee that every generated animation will preserve physical consistency.

Interactive Web Design: The 24 Solar Terms

#

A second test asked K3.1 Pro to create a standalone HTML page about the 24 Solar Terms.

The specification called for a slowly transitioning Three.js scene in the hero section, followed by 24 sequential sections, each containing a short explanation and a corresponding seasonal color palette. The visual direction emphasized restrained East Asian aesthetics, whitespace, and minimal decoration.

The reported output included animated seasonal transitions and palette changes linked to scroll position, with generous spacing throughout the page.

The important result is not simply that the model generated HTML and JavaScript. It attempted to translate a subjective visual brief into a cohesive design, combining layout, animation, typography, and interaction in one deliverable.

Godot: A Playable 2.5D Platformer

#

Another test tasked K3.1 Pro with creating a small 2.5D side-scrolling game in Godot 4.

The game featured a cat navigating rainy urban rooftops, collecting lanterns and avoiding patrol drones. The requested implementation included a start screen, score tracking, and restart behavior after failure.

The difficult parts involved much more than generating a playable scene. The project required camera-depth tuning, jump physics, collision behavior, drone movement, UI references, font loading, and state cleanup after restarting.

The reported workflow included automated playtesting and more than 20 render checks. The project was packaged so it could be extracted and launched from a clean archive.

These details illustrate the distinction between a generated prototype and a runnable software artifact. Packaging, testing, and recovery from repeated execution are important parts of evaluating an autonomous coding model.

🧊 3D Creation: Blender, FreeCAD and Unreal Engine

#

Kimi K3.1 Pro also emphasizes spatial reasoning and 3D asset creation.

On BenchCAD, Moonshot reported a score of 92.7% with tools, compared with 96.2% for Claude Opus 5.5 and 95.9% for GPT-6 Astra. Without tools, K3.1 Pro scored 72.8%, close to Opus 5.5’s reported 73.0%.

The tool-enabled and tool-free figures measure different configurations. They suggest that external execution capabilities can contribute substantially to the model’s 3D performance.

Parametric Modeling and Script-Based Editing

#

The model can use code to construct and revise geometry in applications such as Blender and FreeCAD. This allows it to create editable scene elements rather than returning only a static rendered image.

For example, a model can be represented as a hierarchy of components with independent geometry, materials, and transformations. The agent can then update a component through a script, render the scene, inspect the result, and make further adjustments.

This is especially relevant to parametric design, where dimensions and relationships can be encoded explicitly and revised without manually rebuilding the entire object.

Moonshot has also introduced 3D Make, a feature that supports online 3D model preview and editing, with options to download models for 3D printing or export them as videos and Live Photos.

Architectural Reconstruction in Blender

#

One reported demonstration tasked K3.1 Pro with reconstructing the Hall of Prayer for Good Harvests at Beijing’s Temple of Heaven.

The workflow involved researching reference dimensions, generating Blender Python scripts, and constructing the major architectural elements: a triple-tiered marble base, blue glazed-tile roofs, a gilded dome, and red pillars.

The model also added trees and scene lighting, then performed multiple render-and-critique iterations. The delivered artifacts reportedly included editable Blender project files, scripts, and a web-viewable presentation.

A related task reconstructed a simplified Palace of Fine Arts in San Francisco using reference photographs, with the columns, rotunda, dome, and lighting represented as separate editable objects.

These cases show the potential of an agent that can combine research, code generation, scene construction, and visual inspection in one workflow. The final quality would still need to be assessed against the accuracy of the reference geometry and the requirements of a production asset pipeline.

More Ambitious 3D Demonstrations

#

Other reported examples span several types of technical content:

• Unreal Engine: A cloud-palace environment with a reported radius of 1,750 meters, 36 colonnade bays, and approximately 36 million source triangles.
• Godot: A scripted 3D action-adventure prototype with five explorable regions, two-phase boss fights, and a day-night cycle.
• FreeCAD: A turbofan engine assembly containing approximately 2,100 solid parts, organized for step-by-step assembly.

The number of generated components or triangles indicates scene complexity, but it does not by itself prove geometric accuracy, mechanical validity, or real-time performance. The relevant evaluation is whether the assets are editable, internally consistent, and suitable for the intended workflow.

🧬 Beyond Code: Debugging a Human Walking Simulator

#

One of Moonshot’s more technically ambitious internal demonstrations used the open-source human-atlas anatomical model to build a human walking dynamics simulator.

The model contains 296 skeletal components, including cartilage, and 402 muscle components. The task required K3.1 Pro to work with the simulation environment, analyze failures, refine the controller, and evaluate its results.

According to Moonshot, the evaluation involved 167 tasks and 420 independent review rounds. Early attempts failed within seconds, but iterative debugging eventually produced continuous alternating strides lasting 11.2 seconds.

Diagnosing the Underlying Physics

#

The most revealing part of the demonstration occurred when adjustments to the controller’s reward function failed to improve arm motion.

Instead of continuing to tune the same parameters, K3.1 Pro reportedly performed open-loop probes of the simulation dynamics. It identified 36 errors in shoulder-muscle moment-arm calculations, corrected the underlying physics, and resumed controller training.

This is a different form of problem-solving from ordinary code repair.

A software agent may be able to find a syntax error or failing test by inspecting source files. A dynamics simulator introduces another level of complexity: the code may execute correctly while the modeled physical relationships remain wrong.

Diagnosing such a failure requires connecting the observed behavior with the mathematical and mechanical assumptions underlying the simulation.

The demonstration is an internal result rather than an independent evaluation of general physical reasoning, but it illustrates the direction of Moonshot’s long-horizon training objectives.

💰 Financial Research and Reliability

#

Kimi K3.1 Pro also targets workflows where accuracy, source verification, and reproducible calculations matter more than presentation quality alone.

Financial analysis typically requires an agent to connect source documents, preserve numerical consistency, construct linked formulas, and produce conclusions that can be audited.

Moonshot reports improvements across financial modeling, investment research, and business-operation tasks.

OfficeQA Pro v2 and Frontier Finance

#

On OfficeQA Pro v2, K3.1 Pro reportedly scored 64.4, ahead of Claude Opus 5.5’s 62.2 in the cited comparison.

The benchmark evaluates financial-document analysis involving retrieval, cross-document verification, and multi-step numerical reasoning.

On Frontier Finance, which evaluates workflows ranging from target screening to profitability analysis, K3.1 Pro reportedly scored 58.6.

These results suggest that the model can coordinate research, calculations, and analysis within a longer business workflow. They do not guarantee that every generated spreadsheet or investment report will be correct.

Hallucination Reduction

#

Moonshot’s internal Kimi Hallu Bench Hard stress test reported that K3.1 Pro reduced its claim-level hallucination rate from 23.4% for K3 to 8.3%.

In the same reported hard-with-tools evaluation, Claude Opus 5.5 scored 15.0%, while GPT-6 Astra scored 5.5%.

These figures are from a deliberately challenging internal evaluation. They should not be interpreted as the everyday probability that a model will make a factual error in ordinary use.

Separately, Artificial Analysis reported that the hallucination rate measured by its own AA-Omniscience evaluation fell from 53% for K3 to 44% for K3.1 Pro. These results are directionally consistent but use a different methodology and should not be combined into a single error-rate estimate.

For financial and enterprise applications, the practical requirement remains the same: important claims should be traceable to original documents, calculations should be checked, and outputs should be evaluated against the requirements of the task.

🔐 Open Weights, Cybersecurity and Deployment

#

At launch, Kimi K3.1 Pro is available through Moonshot’s hosted services and API, while the complete model weights remain unpublished.

Moonshot has committed to releasing the weights by November 4, 2026. As of October 25, developers should therefore distinguish between using the hosted model and downloading the full weights for self-managed deployment.

Cybersecurity Safeguards Before the Weight Release

#

Moonshot has also described changes intended to reduce the model’s offensive cybersecurity capabilities before publishing the weights.

According to the company, it removed cyber-related data from parts of the training corpus and filtered relevant tasks from its reinforcement-learning environments while preserving general-purpose coding abilities.

On Moonshot’s internal xdev-bench-hard vulnerability research benchmark, the resulting model scored 81.0%, compared with 90.5% for a similar checkpoint without those risk-reduction measures.

These are company-reported results, not independent confirmation that the measures prevent misuse. The distinction matters because distributing model weights makes the model easier to run and modify outside the provider’s control.

Reuters also reported on Moonshot’s cybersecurity risk-reduction work and the challenges associated with releasing highly capable weights.

Where Kimi K3.1 Pro Is Available

#

The model is currently available through:

• Kimi Web and mobile apps: Hosted access through Kimi.com and supported mobile platforms.
• Kimi Work: Desktop productivity and research workflows.
• Kimi Code: Coding-agent workflows and model selection within the development environment.
• Kimi API: Programmatic integration using kimi-k3.1-pro or kimi-k3.1-pro-highspeed.
• Enterprise deployment options: Organization-level controls and zero-data-retention arrangements for eligible API customers.

The main difference between hosted access and a future self-managed deployment is the level of operational control. Hosted services reduce infrastructure and maintenance requirements, whereas open weights can allow organizations to deploy the model on systems they control, subject to the eventual license and hardware requirements.

The weights’ publication date and license will therefore be important factors for developers evaluating on-premises deployment, fine-tuning, and data governance.

💵 API Pricing: Lower Rates Than Kimi K3

#

Kimi K3.1 Pro’s API pricing is another part of its competitive positioning.

Moonshot’s listed standard rates are $2 per million input tokens and $10 per million output tokens. Cached input tokens are priced at $0.20 per million tokens.

The Highspeed variant is priced at $4 per million input tokens and $20 per million output tokens, with cached input priced at the same $0.20 per million tokens.

Pricing Comparison

#

The standard K3.1 Pro rates are approximately one-third lower than the previous K3 list prices shown above, while the Highspeed variant doubles the standard model’s input and output rates.

Moonshot says Highspeed can provide two to three times the output speed of the standard model. Actual throughput depends on serving conditions, context length, concurrency, and workload characteristics; the speed claim should not be interpreted as a guarantee for every request.

Cost per Task Matters More Than Token Price Alone

#

Low token prices do not necessarily translate into the lowest total cost for a completed task.

An agentic workflow may consume substantial reasoning and output tokens while repeatedly inspecting files, running tools, and correcting errors. A model that costs less per token but requires many more attempts can be more expensive than a model that finishes quickly.

Artificial Analysis estimated K3.1 Pro’s cost at approximately $2.11 per Intelligence Index task under its evaluation methodology. This is a workload-specific estimate, not a fixed price for every API request.

For developers, meaningful cost comparisons should consider task success rate, total tokens consumed, output verbosity, latency, tool calls, and the amount of human correction required.

📝 Conclusion: A Stronger Model for Long-Horizon Work

#

Kimi K3.1 Pro represents a significant upgrade over Kimi K3, particularly in agentic coding, workflow automation, 3D creation, and financial research.

Its reported 2.8-trillion-parameter architecture, 104-billion active parameter count, and one-million-token context window provide the foundation for complex workloads. More importantly, Moonshot’s emphasis on large-scale reinforcement learning targets the execution and debugging skills required to turn model outputs into working applications and technical artifacts.

The reported benchmark results are competitive with leading proprietary models, although the evidence varies by evaluation. Artificial Analysis places K3.1 Pro fourth on its Intelligence Index snapshot, while the strongest coding and workflow gains are supported by a mixture of third-party and vendor-reported tests.

The SVG, web, game-development, Blender, FreeCAD, and human-simulation demonstrations further illustrate the model’s ambition to perform complete work rather than stop at code generation. These capabilities will need continued independent testing to establish their reliability across a broad range of real-world projects.

As of October 25, 2026, hosted access is available and API pricing is competitive, but the full weights have not yet been released. The planned November 4 release will be an important milestone for organizations interested in local deployment, fine-tuning, and deeper control over their AI infrastructure.

The next test for Kimi K3.1 Pro is reproducibility. If independent developers can consistently reproduce its claimed gains in long-horizon coding, tool use, financial analysis, and 3D creation, Moonshot’s reinforcement-learning strategy could become an important example of how frontier AI capability can improve through more effective post-training—not only through larger pretrained models.