Mozilla's State of Open Source AI report finds that the performance gap between leading US closed AI models and the best Chinese open-weight models has narrowed to just 4.4 months. Ars Technica's coverage of the report highlights that, for many organizations, choosing a costly closed model by default may no longer be justified for routine tasks.
Key measured results
- According to Mozilla, Moonshot AI's open-weight Kimi K3 trails Anthropic's closed Fable 5 by only three points on the Artificial Analysis Intelligence Index, while Kimi's cost is roughly 30 percent of the rival's.
- Independent tests by Vals AI show the Chinese Zhipu AI open-weight GLM 5.2 performed within one point of Anthropic's Claude Opus 4.7 and 4.8, at one-fifth the per-task cost.
- On the OpenRouter AI marketplace in August 2026, eight of the ten most popular models by token volume were open-weight models, indicating increased usage of open models.
Where closed models still hold an edge
The report cites a METR nonprofit research group's "time horizon" metric, which measures the length of tasks (in human-expert time) that a model can solve with at least 50 percent success. Current figures indicate closed models reliably handle tasks roughly 1.7 times longer than open models. Practically, this means closed models retain advantages for 8–12 hour workflows: tasks shorter than about eight hours are often within reach of open models, while tasks longer than 12 hours are not reliably solved by either model type.
Raffi Krikorian, Mozilla's Chief Technology Officer, says the price premium for closed models is justified only for specific workloads—such as expert-level professional tasks, applications that require intensive information retrieval, or operations demanding long context windows.
Practical use cases and market effects
- DoorDash, for example, uses the Kimi model for routine tasks and reserves Fable for more complex, time-consuming work that even human experts find demanding.
- Despite growing adoption, closed models continue to dominate revenue: a Linux Foundation study found open models accounted for only 4 percent of total revenue during the period from May to September 2025.
Strategy and geopolitical notes
The report observes that the most widely used open models globally are currently largely developed in China. Chinese labs are following a strategy similar to what American companies did with Android: they make the model freely available while retaining control over the surrounding ecosystem.
Krikorian warned this dynamic brings both diversity and market concentration, and he urged American and European developers to enter the open competition to prevent a single country from dictating global defaults.
Conclusion
Current measurements show Chinese open-weight models are rapidly approaching Western closed-model performance while offering substantial cost advantages on many tasks. Open models are now competitive for shorter, routine workflows, but closed solutions still hold advantages for longer, specialized, or expert-level tasks and remain dominant on the revenue side.
Editor’s note: an AI assistant helped prepare this article; the final content was edited and verified by our journalist.
