Open-source AI models close the gap with pricey frontier alternatives

The performance disparity between closed, proprietary AI models from Silicon Valley and openly available models from Chinese competitors has narrowed considerably. New research finds that frontier models from companies like Anthropic now offer only a marginal advantage over publicly available alternatives, raising questions about when—and whether—organizations should pay a premium for closed systems.

The narrowing lead#

Mozilla's latest State of Open Source AI report, released in mid-September, measures this gap at approximately 4.4 months of development difference. The comparison uses the Artificial Analysis Intelligence Index as a benchmark. A particularly striking example involves Moonshot AI's open Kimi K3 model, which trails Anthropic's proprietary Fable 5 by only three performance points on the index—yet costs just 30 percent as much.

When proprietary actually matters#

Despite the closing gap, organizations continue to adopt expensive closed models for specific scenarios. According to Mozilla's chief technology officer, Raffi Krikorian, the real value of paid frontier models lies in three narrow categories: expert-level professional work, complex information retrieval tasks, and handling unusually long documents or conversations.

The choice to invest in proprietary systems, he emphasizes, tends to be driven by particular use cases rather than an organization-wide strategy.

Why cost and control matter#

The distinction between these model types hinges on more than just performance numbers. Open-weights systems allow developers to download and run the core model components independently, though creators typically keep proprietary information—such as training methods, data sources, and processing pipelines—under wraps.

Proprietary alternatives from US companies like OpenAI and Anthropic require users to access models through APIs or hosted platforms, but they arrive with built-in advantages: immediate usability, formal compliance frameworks, professional support, and legal accountability. For many organizations, especially those without dedicated teams to optimize open models, this turnkey approach justifies the higher price tag.