Open Weights and Chinese Models Narrowed the Frontier Gap to About 4.4 Months

According to an exclusive by Ars Technica, a Mozilla report published September 15—its "State of Open Source AI"—finds that the performance gap between US closed frontier models and the best Chinese open-weight models has narrowed to just about 4.4 months.

The numbers#

Per Ars, Moonshot's open model Kimi K3 scores a composite that is just three points behind Anthropic's closed frontier model Fable 5 on the Artificial Analysis Intelligence Index, while costing only about 30 percent as much. Mozilla CTO Raffi Krikorian told Ars that a closed model "earns its premium" in "expert professional work, high-intensity retrieval, and long context," and that the decision to pay is "workload-specific rather than organization-specific."

Where "worth it" ends#

Ars cites the nonprofit METR's "time horizon" measure: the best closed model can reliably do a job about 1.7 times as long as the best open model can. Krikorian's example: if the open frontier handles a seven-hour job while the closed one handles twelve, then in four months the open model handles twelve too. In short, paying for a closed model buys roughly a four-month head start at about five times the per-task cost, and only for tasks in the eight-to-twelve-hour band. Benchmarking firm Vals AI, re-running models on a single neutral harness, found that China's Z.ai (Zhipu) open GLM 5.2 scored within a point of Anthropic's Claude Opus 4.7 and 4.8 on Terminal-Bench 2.1 while costing about five times less per task.

Concentration of power#

Ars notes that eight of OpenRouter's top ten models by August token volume provide open weights, yet a Linux Foundation study by Frank Nagle and Daniel Yue found open models earned just 4 percent of overall revenue (May-September 2025 data). Krikorian warns that "most of the open models the world runs on are Chinese," calling it both a plurality and a concentration, and urged US and European labs to enter "the same open lane, so that no single country sets the world's defaults."

Our take: A practical read for buyers: hand routine tasks under eight hours to cheap open models, and pay for closed only on high-value 8-to-12-hour work with a deadline that lands before the open frontier catches up. Past twelve hours, neither class is reliable yet.