K3.1 Pro Technical Capabilities: Code Generation and Long-Horizon Agentic Tasks

1. Model Positioning and Upgrade Priorities

Two keywords define this version update: code capability and long-horizon agents. Compared with the previous generation, the new model does not simply make incremental improvements to general conversational experience. Instead, it concentrates resources on scenarios closer to real-world development and business delivery: the model needs to maintain stable planning and execution across long, multi-step, cross-tool task chains.

According to officially disclosed data, this generation has reached or approached the level of the strongest closed-source models on code, agent, and knowledge benchmarks. More notably, it shows a trend of approaching or even surpassing leading closed-source models in more demanding evaluations, particularly in terminal-constrained tasks and long-horizon code generation—areas where large models have generally performed inconsistently.

2. Terminal-Constrained Tasks: From “Can Write Code” to “Can Operate Real Environments”

In the past, many code model evaluations focused on static code completion or short function generation. While these tasks reflect a model’s grasp of syntax and common patterns, they struggle to measure usability in real terminal environments. Real terminal tasks often have the following characteristics:

• The environment state is unknown, requiring the model to explore before making decisions;
• Command execution may fail, requiring the model to dynamically adjust based on error output;
• Task chains are long, requiring the model to maintain goal consistency across multiple rounds of interaction;
• Operations target real file systems, processes, and tools rather than predefined sandbox interfaces.

The new model scored 65.7% on Terminal-Bench 4.0, higher than GPT-6 Astra’s 58.2% and only slightly behind Claude Opus 5.5’s 66.4%. This indicates that the model has acquired the basic ability to complete constrained tasks in real terminal environments, not just generate command sequences that “look correct.”

On the internal long-horizon code benchmark K3 Code Bench 3.0 Long-horizon, the score reached 61.2%, a notable improvement over the previous generation’s 55.95%. The core value of this benchmark lies in examining the model’s performance in long code generation, cross-file modifications, and context retention—capabilities that are more critical for automated development of large projects.

3. Game Engines and Parametric Modeling: Extending Tool-Calling Capabilities

The model’s ability to call Unity, Unreal Engine, Godot, and other game engines demonstrates its operability across complex toolchains. Game engines typically involve a large number of APIs, scene management logic, and resource organization methods. The model needs to understand the engine’s basic architecture to generate runnable scripts. In the BenchCAD tool-assisted scenario, the model scored 92.7%, approaching the level of leading closed-source models, indicating it can proficiently drive Blender and CAD scripts to complete parametric 3D modeling.

The practical value of these capabilities: developers can describe modeling requirements in natural language, and the model generates executable parametric scripts, transforming repetitive modeling operations into reusable, adjustable code assets. Compared with manual operations, this approach is better suited for batch variant generation, rapid design validation, or integration with automated pipelines.

4. Financial Workflows: Validating End-to-End Delivery Capability

Finance is a typical application scenario for long-horizon agents. A complete industry research report or financial model often requires data acquisition, metric calculation, model construction, text writing, and cross-validation across multiple stages. If the model can only complete one stage, users still need to manually connect multiple tools, resulting in limited efficiency gains.

The new model scored 64.4 on the external benchmark OfficeQA Pro v2, which is claimed to be the current best result. This benchmark evaluates the model’s comprehensive performance on office and financial tasks, indicating that the model can deliver financial modeling, investment analysis, and industry research workflows end-to-end. For finance professionals who handle large volumes of structured data and unstructured text, this capability means they can focus more energy on judgment and decision-making rather than repetitive data processing and document writing.

5. Factual Consistency and Safety Governance: Reliability as a Prerequisite for Engineering

In long-horizon tasks, factual errors are often more destructive than in short Q&A scenarios. Errors introduced early in a task propagate along the task chain, ultimately making the entire deliverable unusable. Therefore, improving factual consistency is a key step in enhancing long-horizon agent usability.

In the internal stress test Kimi Hallu Bench Hard, the error rate for individual factual statements dropped from 23.4% in the previous generation to 8.3%, and the proportion of responses containing at least one factual error was only about 42.4% of the predecessor. The official announcement also emphasizes that this test deliberately focuses on error-prone situations, and the error rate does not represent overall performance in daily use. Even so, this data still demonstrates a significant improvement in the model’s reliability in confusing, high-difficulty knowledge scenarios.

On safety governance, the official team identified and removed network attack and defense-related data during both mid-training and reinforcement learning stages to reduce abuse risk. On the self-developed xdev-bench-hard, the unfiltered version with pass@3 can solve 90.5% of problems, while the final release version solves 81.0%. This gap reflects the trade-off between safety filtering and capability retention: the model retains most of its ability to solve complex development tasks while reducing potential risks. For enterprise users, this predictable safety boundary is more practically meaningful than simply pursuing the highest score.

6. Pricing Strategy and Cost Optimization

In terms of pricing, the new generation has been reduced compared with the previous generation. The standard version is priced at 13 yuan per million tokens for input and 68 yuan per million tokens for output; the high-speed version is priced at 2x the standard version, at 26 yuan input and 136 yuan output. The context cache hit price is uniformly 1.3 yuan per million tokens, a design that encourages users to prioritize cache utilization in long conversations or long document processing scenarios, thereby reducing repeated computation costs.

Cache write pricing is divided into two tiers: 5-minute write and 1-hour write. For the standard version, these are 13 yuan and 26 yuan per million tokens respectively; for the high-speed version, 26 yuan and 52 yuan. This tiered pricing suits scenarios with different latency requirements: for batch tasks that frequently reuse context, choosing a longer cache period can effectively amortize costs; for interactive development scenarios with higher real-time requirements, short-period caching or the high-speed version can be selected.

From a cost structure perspective, the new model has reduced both input and output prices while maintaining the same low cache hit pricing strategy as the previous generation. For teams with high-frequency calls for code generation, long document analysis, or agent tasks, the cumulative cost savings from this adjustment could be considerable.

7. Engineering Deployment Recommendations

For developers considering incorporating this model into their workflows, the following points merit attention:

1. Validate terminal tasks in controlled environments first: The model’s performance in real terminal environments still differs from static code generation. It is recommended to verify task success rates in containers or sandboxes first, then gradually expand permissions.

2. Leverage context caching: Long-horizon agent tasks often require maintaining large amounts of context across multiple rounds of interaction. Designing session structures reasonably and maximizing cache hit rates can significantly reduce inference costs.

3. Retain human review for financial scenarios: Although the model performs well on financial benchmarks, content involving investment decisions and financial disclosures should still have human review checkpoints to ensure compliance and accuracy.

4. Align safety boundaries with enterprise policies: The model’s capability retention rate after safety filtering is approximately 81.0%. Enterprises should evaluate whether this trade-off suits their business needs and, if necessary, supplement it through private deployment or additional access control strategies.

Overall, this update forms a relatively complete capability loop across long-horizon code, terminal operations, tool calling, and knowledge reliability. For teams that need to build complex development automation pipelines or industry-specific agents, this model offers a relatively balanced choice in terms of capability and cost.