Open AI Models Trail Frontier Rivals by Just 4.4 Months at One Fifth the Cost
Mozilla finds Chinese open AI models trail US frontier systems by only 4.4 months, while comparable tasks cost about one fifth as much to complete today.
Summary
Mozilla’s State of Open Source AI, published September 15, says leading US closed frontier systems are only 4.4 months ahead of Chinese open weights models. Moonshot AI’s Kimi K3 trails Anthropic’s Fable 5 by three points on the Artificial Analysis Intelligence Index while costing 30 percent as much; DoorDash uses Kimi routinely and Fable for harder work. Open weights are downloadable but often omit training data, pipelines and code. Anthropic and OpenAI charge more for proprietary, ready made systems with compliance, support and accountability; Mozilla CTO Raffi Krikorian recommends open models by default, reserving closed systems for expert work, intensive retrieval and long context.
METR’s time horizon measures tasks models complete with 50 percent reliability and has doubled at an accelerating cadence. The best closed model handles work 1.7 times longer than the best open model: seven hours versus 12, then, four months later, 12 versus about 20. Both can generally manage tasks below eight hours, closed models lead from eight to 12, and neither reliably exceeds 12 today. Lab built harnesses can skew results, but on Vals AI’s neutral harness in Terminal Bench 2.1, Z.ai’s GLM 5.2 finished within one point of Anthropic’s Claude Opus 4.7 and 4.8 at about one fifth the per task cost.
Open model use has accelerated since Mozilla’s July 14 inaugural report; eight of OpenRouter’s top 10 models by August 2026 token volume had open weights. Yet Linux Foundation research by Frank Nagle and Daniel Yue assigned open models 4 percent of revenue versus 96 percent for closed models from May through September 2025, before the latest surge. With top open models concentrated in China and closed leaders in the US, Krikorian wants US and European labs, public compute programs, neutral foundations, beneficiary companies and philanthropy to fund fully open models, standards, evaluations and audits; Switzerland’s Apertus is his example.
Positives
- Kimi K3 delivers a score within three points of Fable 5 while costing only 30 percent as much.
- GLM 5.2 matched Claude Opus 4.7 and 4.8 within one point at roughly one fifth the completed task cost.
- Eight of OpenRouter’s top 10 models by August 2026 token volume offered open weights.
- DoorDash cuts routine AI costs with Kimi while retaining Fable for work demanding greater capability.
- Tasks below eight hours can generally move to cheaper open models without sacrificing access to adequate capability.
Risks & concerns
- Closed models retain a meaningful advantage on eight to 12 hour tasks, particularly when deadlines arrive before open models catch up.
- Neither open nor closed models generally complete tasks exceeding 12 human expert hours with 50 percent reliability.
- Open weights frequently exclude training data, pipelines and code, limiting transparency and independent scrutiny.
- Open models captured only 4 percent of revenue in May through September 2025, versus 96 percent for closed systems.
- China’s concentration of leading open models risks allowing one country’s ecosystem to shape global AI defaults.