Open AI models trail closed ones by just 4 months — but cost a fifth as much

The capability gap between the frontier closed AI models of American companies and the best open-weight models from China has narrowed to 4.4 months. So says a new Mozilla report, “State of Open Source AI,” published on September 15 and shared in advance with Ars Technica. The report's conclusion is simple: most organizations should use open models by default for the bulk of their tasks.

The numbers bear this out. Moonshot AI's open Kimi K3 model trails Anthropic's closed Fable 5 by just three points on the composite Artificial Analysis Intelligence Index, while costing 30% as much. Mozilla CTO Raffi Krikorian frames the principle this way: paying for a closed model makes sense only for a narrow class of tasks — expert professional work, intensive search, and long context. That decision, he says, depends on the specific workload, not on the organization as a whole.

Where the line falls

The researchers measure the gap through a “time horizon” — the length of a task (by the time a human expert spends on it) that a model completes with 50% reliability. The best closed model handles a task 1.7 times longer than the best open one. “If the open frontier takes on a seven-hour job, the closed one takes on a twelve-hour job,” Krikorian explains. “Four months from now the open model will manage twelve hours, and the closed one will be up to around twenty.” Tasks shorter than eight hours are within reach of any model; those longer than twelve, of none so far.

To level the playing field, Vals AI runs the models through its own neutral harness. On the Terminal-Bench 2.1 test, the open GLM 5.2 from China's Z.ai trailed Claude Opus 4.7 and 4.8 by less than a point, at roughly a fifth of the cost per completed task. Hence the formula: a closed model buys a four-month head start at five times the price — but only on tasks eight to twelve hours long. Krikorian's logic: pay when the head start is worth it (say, a deadline arrives before the open models catch up), but not for the routine work you'll soon be able to do at a fifth of the cost.

Who owns the ecosystem

The popularity of open weights is visible on the OpenRouter marketplace: in August 2026, eight of the ten most sought-after models by token volume had open weights. By revenue, though, open models still take only 4% of the market against 96% for closed ones — though these figures date to May–September 2025.

Most of the world's open models are Chinese. What worries Krikorian is not so much their origin as the concentration: the best open models are clustered in China, the best closed ones in the US. “Chinese labs are repeating the move Americans used to win with Android: give it away for free, but own the ecosystem around it,” he says. As a counterweight, he proposes building, on the Linux model, an “alternative coalition” of institutions “with a mission, not a market”: public-compute programs (the Swiss Apertus is one example), neutral foundations, and philanthropy funding the auditing and evaluation of models.

Source: Ars Technica