Mozilla: open models trail frontier AI by just 4.4 months

Mozilla's report, previewed by Ars before its September 15 publication, lands as the second edition of its State of Open Source AI series, following an inaugural report on July 14. The gap it describes is measured several ways, including METR's "time horizon" — how long a task, in expert human hours, a model can complete with a reliable 50 percent success rate. Krikorian puts the current ratio at 1.7×: if open handles a seven-hour job, closed handles a 12-hour one, and in four months open catches up to 12 while closed reaches around 20. Vals AI's neutral-harness testing on Terminal-Bench 2.1 adds a cost dimension, with Z.ai's GLM 5.2 scoring within a point of Anthropic's Claude Opus 4.7 and 4.8 at about five times less per completed task.

The report pairs that capability convergence with a lopsided business picture. On OpenRouter, eight of the top 10 models by token volume in August 2026 offer open weights, yet a Linux Foundation paper by Frank Nagle and Daniel Yue found open models earning just 4 percent of overall revenue versus 96 percent for closed, based on data from May through September of 2025. Krikorian expects that revenue split to have shifted, and points to the geographic concentration behind today's open ecosystem: most open models the world runs on are Chinese, which he compares to the Android playbook of giving software away while owning the surrounding ecosystem.

Krikorian's proposed answer is a coalition — public compute programs, neutral foundations, companies that benefit from commodity models, and philanthropy — modeled loosely on how open source infrastructure like Linux was funded. The stakes hinge on whether such an "alternative coalition" of mission-driven institutions, rather than frontier labs, can materialize before China's position in open models hardens further.