Open-source AI models close gap with closed-source to just 3.3%, DeepSeek V4 Flash tops API usage

Mozilla's 2026 Open Source AI Status Report shows that the capability gap between open-weight and closed-source AI models has narrowed to an average of 3.3% based on Chatbot Arena data. This gap was 8.04% in January 2024, dropped to 0.5% by August 2024, and widened slightly after DeepSeek-R1's release in February 2025, which briefly matched top US models. The 3.3% average gap mainly lies in reasoning, long-context retrieval, and agent tasks, while coding, instruction following, and general knowledge are nearly equal. GPT-4-level model prices fell from $20 per million tokens in late 2022 to $0.40 in December 2025, a 50-fold drop. In API usage, the top five models on OpenRouter are all open-source: DeepSeek V4 Flash leads with 18.4T tokens, followed by Xiaomi Mimo-V2.5 with 14.9T and Tencent Hy3 Preview with 14.8T. Developer adoption of open models is 79%, but only 51% successfully deploy to production, compared to 63% for closed-source models, with performance, integration, and maintenance being key hurdles.