Open AI Models Close Gap on Frontier AI

The performance difference between top-tier AI models from US companies and leading open-weights models developed in China has shrunk to just 4.4 months. This narrowing gap is prompting many organizations to adopt significantly cheaper open-weights models for routine tasks, as highlighted by Mozilla's State of Open Source AI report published on September 15, 2026.

Moonshot AI's Kimi K3, an open model, now rivals Anthropic's Fable 5, a closed frontier model, in AI performance scores, while costing only 30 percent as much. While closed models still offer an advantage for highly specialized professional work and complex retrieval tasks, the cost-effectiveness of open models is making them the default choice for most organizations.

This shift is accelerating the adoption of open models. For instance, DoorDash uses Kimi for daily operations and reserves more advanced models for intricate tasks. The trend indicates a move towards leveraging open-source AI's growing capabilities and economic benefits.

Currently, the best closed models can handle tasks 1.7 times longer than the best open models. However, this advantage is diminishing. Tasks taking less than eight hours are reliably completed by both types of models, making open models a viable and cost-efficient option.

The open-weights AI ecosystem is largely dominated by Chinese companies, with eight of the top 10 models by token volume in August 2026 being open-weights. This concentration raises concerns about a single country setting global AI defaults.

Mozilla advocates for a collaborative approach, akin to the development of open-source software like Linux, to foster competition and prevent such concentration. This would involve public compute programs, neutral foundations, and industry participation to ensure a pluralistic AI future.