Moonshot Cuts AI Inference Costs by Third as Kimi K3.1 Pro Narrows Market Gap

Token prices for Moonshot AI’s latest inference services have dropped by 33 percent, a strategic move that coincides with the release of Kimi K3.1 Pro and positions the Chinese firm to capture more market share in the high-stakes AI sector. This price adjustment brings input costs down to $2 per million tokens and output costs to $10 per million tokens, signaling a shift toward greater affordability for enterprise clients who are increasingly sensitive to operational expenditure.

The performance gains accompanying this price cut are substantial. Kimi K3.1 Pro achieved a score of 55 on the Artificial Analysis Intelligence Index, marking an 11-point increase over its predecessor. This result places the model in fourth position globally, trailing only Claude Opus 5.5, Claude Sonnet 5.5, and Muse Wonder Full. The proximity to the top tier is notable, as the new model sits just three points behind the highest-scoring proprietary systems.

From a market dynamics perspective, the upcoming release of open model weights is a critical variable. Moonshot has scheduled the release of these weights for November 4, 2026. If this timeline holds, Kimi K3.1 Pro will become the highest-scoring open-weights model available. This development would compress the performance gap between open and proprietary architectures from 12 points to just three, potentially altering the cost-benefit calculations for companies choosing between custom-built and off-the-shelf AI solutions.

The model’s improvements are most visible in agentic tasks, where efficiency directly impacts workflow automation. On the GDPval-AA benchmark, the new model reached an Elo rating of 1,846, a 312-point surge over the previous version. In terminal-based coding evaluations, it recorded a 65.7 percent success rate, narrowly missing the 66.4 percent achieved by Claude Opus 5.5 while significantly outperforming GPT-6 Astra. These metrics suggest a maturing capability in executing complex, multi-step software tasks.

Reliability has also improved, with internal tests showing a reduction in hallucination rates. Claim-level hallucinations dropped from 23.4 percent to 8.3 percent on Moonshot’s internal stress tests. While this remains higher than the 5.5 percent recorded by GPT-6 Astra, the trajectory indicates a meaningful step forward in factual consistency. In financial applications, the model gained 11.6 points in research tasks and 7.2 points in modeling, areas where precision is essential for mid-Atlantic financial institutions.

The technical architecture retains the same parameter count as its predecessor, with 2.8 trillion total parameters and 104 billion active parameters. The model supports a context window of one million tokens and processes both text and image data. According to Artificial Analysis, it currently sits at the top of the Pareto frontier for models of its size, meaning no other model with a comparable parameter count scores higher on the intelligence index. This efficiency ratio is a key selling point for infrastructure-heavy deployments.

Moonshot has also emphasized the model’s capabilities in 3D modeling and game engine integration, supporting direct work with Unity, Unreal Engine, and Godot. On the BenchCAD benchmark, the system achieved a 92.7 percent score when using tools, approaching the performance of leading competitors. The company has launched a "3D Make" feature that allows users to preview, edit, and export models, catering to industries that rely on rapid prototyping and architectural visualization.

In a move to balance capability with safety for the open-weights release, Moonshot removed specific cybersecurity data from its training corpus. This decision reduced the model’s ability to solve advanced vulnerability tasks from 90.5 percent to 81.0 percent on internal hard benchmarks. The company argues that while open models should not push the absolute frontier of offensive cyber capabilities, they must still evolve to support defensive auditing and legacy system maintenance, a stance that reflects the broader industry debate on AI security.