The Mozilla Foundation stated in the latest version of its "State of Open-Source AI" report that the performance gap between frontier models of American tech companies and the best weight-open models from Chinese companies has compressed to only 4.4 months. The report, published on September 15, explains why an increasing number of companies have turned to open and much cheaper models for their everyday tasks and identifies a narrow range of workloads for which spending on frontier models remains justified.
Mozilla's central recommendation is explicit: most organizations should consider open models as the default option for the bulk of their work.
Kimi K3 only three points behind
The most notable example the report cites is the flagship weight-open model Kimi K3 from the Chinese company Moonshot AI. According to the Artificial Analysis Intelligence Index, this model stands only three points behind Anthropic's closed frontier model Fable 5, while costing less than 30 percent of it. Kimi K3 also made its debut in front-end coding on the LMArena platform with a score of 1679, leading in six of the seven domains of that test, as well as in coding, instruction-following, and general knowledge.
Rafi Krikorian, Mozilla's chief technology officer, explained when paying for closed models is sensible, saying these models hold high value in several respects: professional specialized work, high-intensity retrieval, and long-context processing. He emphasized that the decision to pay for closed models should be based on the type of workload, not on the nature of the organization itself.
The difference between "open-source" and "weight-open"
Mozilla's report points to an important technical distinction. Weight-open models allow any user to download the model's core components and run them on their own hardware, but developers typically do not release vital information such as training data, data pipelines, and training code. In contrast, American companies like Anthropic and OpenAI chiefly offer proprietary, fully closed models that require payment for access.
Still, closed models retain enterprise customers because they are offered as ready-made services and come with adaptation, support, and accountability packages. According to Krikorian, many organizations simply lack the personnel needed to effectively operate open models - a point previous Mozilla findings also confirm: 79 percent of developers use open models, but only 51 percent of them have deployed those models to production, while the figure for closed models is 63 percent. Alvaro Ruiz Cobre from SlashData, which conducted the report's survey, views this gap as a sign of infrastructure shortage, not model quality weakness.
Two-tier strategy of companies
The mixed-use pattern is becoming the common practice. For example, food-delivery company DoorDash uses Kimi for routine tasks and resorts to the Fable model for more complex duties that typically take human specialists longer to complete.
Measuring the gap: task time horizon
The narrowing gap between the two model categories is measured by various metrics. The nonprofit research organization METR defined a model's "time horizon": the length of tasks - measured by the time required of human experts - that the model can perform with a reliable 50 percent success rate. This time horizon has been doubling regularly at an accelerating rate.
Accordingly, the best current closed model can reliably perform tasks 1.7 times longer than the best open model. Epoch Institute's capabilities index also estimated the average gap on September 1, 2026, at 8 points, equivalent to about four months. The report's summary is clear: closed models handle eight- to twelve-hour tasks and open models reach the same point four months later; below eight hours both succeed and above twelve hours neither do.
Testing on even footing
To eliminate the effect of infrastructure differences, benchmarking company Vals AI evaluated open and closed models on its neutral testing platform. In the Terminal-Bench 2.1 test, the weight-open GLM 5.2 from Chinese company Z.ai (Zhipu AI) achieved results comparable to Anthropic's Claude Opus 4.7 and 4.8, at about five times lower cost per completed task.
In other words, paying for closed frontier models effectively buys an almost four-month "head start" at more than five times the cost - and this advantage only makes sense in current conditions and solely for tasks that take eight to twelve hours.
Krikorian says that when finishing earlier truly matters - for example, when facing a deadline - rushing makes sense; but routine work that will continue next quarter does not justify that cost, because the same work can soon be done at one-fifth the cost.
The bigger picture: economics and geopolitics
The report is an updated version of Mozilla's initial report released on July 14, which has since helped accelerate the adoption of open models. Findings from that version showed inference costs at the GPT-4 level fell from about $20 per million tokens three years ago to about $0.40 - a 50-fold decrease that Mozilla largely attributes to competition from weight-open models.
But what Mozilla calls the "bitter truth" is the gap between usage and revenue: open models provide about one-third of actual AI usage but account for only 4 percent of the industry's revenue.
On the geopolitical front, the balance is shifting. Open-source AI adoption rates in China and East Asia have reached 89 percent, and the share of Chinese weight-open models in OpenRouter's tokens rose from under 2 percent in late 2024 to over 45 percent of weekly traffic by April 2026. Mozilla attributes this trend to deliberate state policy, including the State Council of China's "AI Plus" initiative. The report concludes that the top four frontier models remain closed - but the next four models are open.
Details of the report were provided to Ars Technica ahead of public release and the full version is available on the report's dedicated website.