Open-weight AI models have captured the majority of production tokens routed through OpenRouter, marking a decisive shift in how developers deploy AI systems, according to Mozilla's comprehensive State of Open Source AI report.

The top seven models by token volume on the platform now all ship open weights, with closed models from major labs appearing further down the rankings. This represents a dramatic reversal from late 2025, when open-weight models handled roughly one-third of tokens.

The performance gap between open and closed models has compressed to just 4 points on the Artificial Analysis Intelligence Index. Moonshot AI's Kimi K3 ranks fourth overall with a score of 57, trailing Claude Opus 5's frontier-leading 61 by the smallest margin yet recorded.

Where open models lead and lag

Open models now match or exceed closed alternatives in specific domains. K3 leads OpenAI's models on frontend coding tasks and scores competitively on agentic terminal work, winning three of four benchmark categories.

Closed models retain advantages in professional knowledge work and long-context reliability. Anthropic's Fable 5 leads K3 by 92 Elo points on expert-graded professional tasks, the largest separation among shared benchmarks.

Inference costs have plummeted 50-fold over 36 months, with the frontier price dropping 112 times from GPT-4's $45 launch price. Open models typically cost about one-third the price of comparable closed alternatives.

The report highlights a split between the "open frontier" and "deployable open frontier." While K3's weights are downloadable, serving the model requires at least eight 8-GPU nodes, limiting deployment to well-resourced organizations.

Mozilla CTO Raffi Krikorian framed the shift as reminiscent of the browser wars, when open alternatives prevented any single company from controlling web access. The report notes that corporate heavyweights including Microsoft, Meta, and NVIDIA recently signed an open letter supporting open-weight development.

The token volume data covers July 2026, when K3's API launched mid-month and weights released on July 27.