Mozilla report: open AI models trail frontier labs by just 4.4 months

The performance gap between the leading proprietary AI models built by U.S. tech companies and the best open-weight models — many of them Chinese — has narrowed to just 4.4 months, according to a new Mozilla report published on September 15. The finding suggests that paying for closed frontier AI now buys a shrinking head start at roughly five times the cost per task.

A Closing Window

Mozilla's second edition of its State of Open Source AI report, previewed exclusively by Ars Technica ahead of publication, quantifies what many in the industry have suspected. Moonshot AI's Kimi K3 now scores only three points behind Anthropic's Fable 5 on the Artificial Analysis Intelligence Index — at about 30 percent of the cost. Mozilla CTO Raffi Krikorian framed the tradeoff as workload-specific: closed models earn their premium mainly on expert professional tasks requiring 8 to 12 hours, while open models handle routine work at a fraction of the price.arstechnica+3

The shift is already reshaping enterprise behavior. DoorDash ↘ 0.14% now uses Kimi for routine work while reserving Fable for harder tasks, and eight of the top 10 models by token volume on the routing platform OpenRouter in August 2026 offered open weights. Yet the business picture remains lopsided — a Linux Foundation study found open models capturing just 4 percent of overall AI revenue despite powering a growing share of usage.hellomarvisaitoday+1

A Geopolitical Dimension

The report lands alongside a separate analysis from the Center for Strategic and International Studies, published September 14, which cast China's open-weight model ecosystem as a direct challenge to U.S. AI leadership. CSIS analysts Taylar Rajic and Lauryn Williams wrote that China's approach — forged through export control constraints, indigenous innovation, and what U.S. officials have called "industrial-scale" distillation attacks on American models — now offers a cheaper, more adaptable alternative that is drawing customers away from U.S. frontier labs.csis

Krikorian drew a similar comparison, likening China's open-model strategy to the Android playbook: give the software away while controlling the surrounding ecosystem. His proposed answer is a coalition of public compute programs, neutral foundations, and mission-driven institutions to fund a Western open-weight alternative before China's position hardens further.hellomarvisaitoday

The Stakes for U.S. Policy

The convergence of the Mozilla data and the CSIS analysis sharpens a policy question already under debate in Washington. In July, Nvidia ↗ 2.78% and Microsoft ↘ 0.99% published an open letter establishing the Open Secure AI Alliance for AI Safety, arguing that open-weight models are the future of U.S. AI leadership. Anthropic CEO Dario Amodei responded publicly, expressing skepticism and continuing to advocate for slower development.csis

As Krikorian put it, the decision facing organizations is no longer whether to use open models, but which tasks still justify paying a premium — a calculus that tilts further toward open with each passing month.arstechnica+1