Mozilla Report: Open AI Model Usage Surpasses Closed Models

Mozilla released its "State of Open Source AI" report in July 2026, detailing the performance and usage trends of open-source versus proprietary artificial intelligence models.

The report indicates that in March 2026, the performance gap between open-source and closed AI models averaged 3.3 percent, according to data from Mozilla's Chatbot Arena. This gap has significantly narrowed since previous periods, decreasing from over 8 percent in January 2024 to just 0.5 percent by August 2024.

Following the release of the DeepSeek-R1 model in February 2025, open-source models briefly matched the capabilities of top U.S. proprietary models. By March 2026, the gap widened again to 3.3 percent. However, the report clarifies this is an average, with open models performing comparably to closed models in coding, instruction following, and general knowledge tasks. Shortcomings are more apparent in reasoning, long-context retrieval, and agent-based tasks.

In terms of API usage, open-source models dominated. In June 2026, the top five models by monthly token consumption on the OpenRouter platform were all open-source. DeepSeek V4 Flash led with 18.4 trillion tokens, followed by Xiaomi Mimo-V2.5 and Tencent Hy3 Preview. Meanwhile, the cost per million tokens for GPT-4 level models saw a substantial decrease, from $20 in late 2022 to $0.40 by December 2025, a 50-fold reduction over 36 months.

Developer adoption shows open models are widely used, with 79% of developers utilizing them at parity with closed models. However, deployment into production environments saw higher success rates for closed models (63%) compared to open ones (51%). The report attributes this disparity not to model capabilities, but to operational challenges such as performance, integration, and maintenance.