The capability gap between leading closed AI models from US companies and the strongest Chinese open-weight alternatives has narrowed to an estimated 4.4 months, according to Mozilla’s latest State of Open Source AI report.
The report compares models using the Artificial Analysis Intelligence Index. It says Moonshot AI’s open-weight Kimi K3 scores three points below Anthropic’s closed Fable 5 while costing 30% as much to use. That combination helps explain why organizations may choose lower-cost open models for routine workloads.
Mozilla chief technology officer Raffi Krikorian told Ars Technica that closed systems still earn their premium in specific areas, including expert professional work, intensive retrieval and long-context tasks. His recommendation is therefore workload-specific: use open models by default for much of an organization’s work, then pay for frontier access where evaluations show a material benefit.
The 4.4-month figure is an estimate derived from a composite benchmark, not a guarantee that the models are interchangeable. Scores can hide important differences in reliability, tool support, deployment requirements and performance on a company’s own data. Buyers still need task-level tests, but the report indicates that a blanket frontier-model premium is becoming harder to justify.