Chinese artificial intelligence startup Moonshot AI has launched Kimi K3.1 Pro, its newest AI model, featuring a one million token context window and improved capabilities across coding, 3D development, financial research, and autonomous AI agents, according to the company's announcement.

On Terminal Bench 4.0, a benchmark that evaluates AI agents on real-world programming tasks, Moonshot said Kimi K3.1 Pro scored 65.7%. That result comes within 0.7 percentage points of Anthropic's Claude Opus 5.5, which scored 66.4%, and surpasses OpenAI's GPT 6 Astra at 58.2%. The 0.7-point spread at the top underscores how closely matched the leading models have become on agentic coding tasks.

The model is built on Kimi K3 with additional training focused on strengthening complex reasoning and the ability to complete tasks that require multiple steps and tools. Moonshot also introduced Kimi K3.1 Pro Highspeed, a variant designed to deliver responses two to three times faster than the standard model. Both versions are available through Kimi.ai, Kimi Work, Kimi Code, and the company's API, with full model weights scheduled for release by November 4.

Gains in Long-Horizon Coding

Moonshot reported that Kimi K3.1 Pro scored 61.2% on Kimi Code Bench 3.0 Long horizon, an internal benchmark measuring an AI agent's ability to complete complex programming tasks involving multiple steps, tool use, and extended problem solving. The result marks an improvement over the 55.95% recorded by its predecessor, Kimi K3.

The company also highlighted improvements in coding across game engines such as Unity, Unreal Engine, and Godot, enabling the model to generate interactive environments, websites, and 3D assets. In one demonstration, Kimi K3.1 Pro developed a walking simulator using an anatomical model containing 296 skeletal components and402 muscle components. Across 167 development tasks and 420 reviews, the model refined the simulator until its virtual character could walk continuously for 11.2 seconds.

The model additionally demonstrated stronger 3D modeling performance, scoring 92.7% on BenchCAD, compared with Claude Opus 5.5 at 96.2% and GPT 6 Astra at 95.9%.

Expanded Focus on Financial Research

Beyond coding, Moonshot is positioning Kimi K3.1 Pro for financial analysis, investment research, and business operations. The company reported a score of 64.4 on OfficeQA Pro v2, an evaluation measuring financial document analysis, information retrieval, and numerical reasoning.

On Moonshot's internal evaluations, Kimi K3.1 Pro improved by 7.2 points in financial modeling and 11.6 points in financial research compared with Kimi K3. The company also reported gains in factual accuracy, with the model's hallucination rate declining from 23.4% to 8.3% on its internal Kimi Hallu Bench Hard evaluation. On the same evaluation using tools, Claude Opus 5.5 recorded a 15% hallucination rate, according to Moonshot's published comparisons. The company cautioned that the benchmark specifically tests difficult scenarios and does not represent hallucination rates during everyday use.

Open Weights Release Planned for November

The launch comes as AI developers continue expanding access to advanced models that can be deployed independently of proprietary cloud platforms. Moonshot said Kimi K3.1 Pro is its most capable model in the 3 trillion parameter class, with full weights expected to become publicly available on November 4. Because several of the reported figures come from Moonshot's own evaluations, the public weights will give outside developers a way to test the model's performance on infrastructure they control.

The company also disclosed changes to its training process designed to reduce cybersecurity risks associated with releasing increasingly capable AI models. Moonshot said it removed certain cybersecurity training data and tasks before evaluating the final model on an internal vulnerability research benchmark. The resulting model scored 81% on the evaluation, compared with 90.5% for a version trained without those restrictions, a 9.5-point gap that reflects the trade-off introduced by the restricted training approach. According to the company, the changes were intended to limit offensive cybersecurity capabilities while preserving general coding performance.

Kimi K3.1 Pro is currently accessible through the company's consumer, enterprise, and developer platforms, with the planned model weights release expected to allow developers to run the model on their own infrastructure. The November 4 release is the next milestone to watch, marking the point at which the model can be deployed and evaluated outside Moonshot's own platforms.