Mozilla’s State of Open Source AI report, published on September 15, said organizations should use open models as the default choice for most of their work. Moonshot AI’s open-weight Kimi K3 model trails Anthropic’s closed Fable 5 model by only three points on the Artificial Analysis Intelligence Index while operating at 30% of its cost. According to Mozilla CTO Raffi Krikorian, closed models offer advantages in specialized professional work, intensive information retrieval, and long-context tasks. They are also preferred because they are ready to use and provide compliance packages, support, and accountability. Although open-weight models can be downloaded, their training data, data pipelines, and training code are generally not shared.
DoorDash uses Kimi for routine work and Fable for more difficult tasks. METR’s “time horizon” data, measured against the time human experts take to complete tasks, shows that the best closed model can handle tasks 1.7 times longer than those the best open model can reliably complete. Krikorian said that while an open model can handle a seven-hour task, a closed model can complete a 12-hour task; four months later, the open model reaches 12 hours, while the closed model reaches about 20 hours. While tasks shorter than eight hours can be assigned to either type of model, the current gap mainly emerges in tasks lasting eight to 12 hours. Overall, no model can reliably perform tasks lasting longer than 12 hours.
Specialized software layers that give models access to tools and memory can affect comparisons. In the Terminal-Bench 2.1 evaluation conducted using Vals AI’s neutral software, Z.ai’s GLM 5.2 model came within one point of Anthropic’s Claude Opus 4.7 and 4.8 models and was approximately five times cheaper per completed task. Accordingly, paying for closed models provides a four-month advantage only for tasks lasting eight to 12 hours, at approximately five times the cost.
As of August 2026, eight of the top 10 models on OpenRouter by token volume are open-weight. By contrast, according to Linux Foundation research based on data from May to September 2025, open models received 4% of revenue, while closed models received 96%. Krikorian said the concentration of the best open models in China and closed models in the US poses a risk. He recommended support from public compute programs, neutral foundations, companies and philanthropic organizations to create an alternative ecosystem in the US and Europe, emphasizing that it is difficult to trust models whose training and evaluation processes are unknown.
Why it matters
This table shows that, for organizations, it makes more sense to choose based on task duration, cost, support and compliance requirements rather than relying on a single model type. While open-weight models reduce cost pressures for routine and short tasks, the advantage of closed models becomes more pronounced in longer tasks requiring expertise; however, the fact that none of the options can reliably complete tasks lasting more than 12 hours underscores the limits of automation. The fact that results may vary depending on custom software layers shows that comparisons should not be based solely on raw performance scores. The concentration of usage in open models and revenue in closed models highlights the distinction between adoption and commercial value, as well as the role of support services. At the same time, the inconsistency in the naming of Anthropic models in the source makes it necessary to verify the comparisons.
Term: open weight
Open weight means that an AI model’s trained parameters are available for download; it is not synonymous with open source, as the training data and code may not be open.