Mozilla's first State of Open Source AI report reveals open models now route the majority of production tokens, cost 50x less than two years ago, and Chinese-built models handle 3x more volume than US ones

The Open-Source AI Race Is Settled. The Deployment Race Isn't.

Open-weight models now route the majority of production AI tokens. The five highest-volume models on OpenRouter are all open weights. On Chatbot Arena, the average gap between open and closed models sits at 3.3%, down from over 8% two years ago 1.

For anyone still asking whether open models can compete, Mozilla's first State of Open Source AI report has a clear answer. That question is settled. The contest moved to a different layer entirely.

Consider the cost collapse. GPT-4-class inference fell from $20 to $0.40 per million tokens over 36 months, a 50-fold drop that Mozilla says outpaced both dotcom-era bandwidth declines and PC-compute price curves 1. Intelligence, as a raw input, is getting cheaper faster than the infrastructure that carried the last tech boom.

The geographic split is just as stark. Chinese-built open models route roughly 18 trillion weekly tokens against about 5.5 trillion for US-built models, a more than 3-to-1 advantage, according to analysis cited in the report 1. Anthropic's closed Claude models are the next US-built entrants on the volume leaderboard, and they sit below the entire open-weight top five 1. One caveat worth noting: by request count, not token volume, closed US providers still lead. Open models dominate on volume because coding and agentic workloads consume enormous token counts per request 1.

The capability picture is nuanced but decisive for most use cases. Open models have reached parity on coding, instruction-following, and general knowledge. The remaining 3.3% gap concentrates in reasoning, long-context retrieval, and agentic tasks 1. In August 2024 the gap briefly collapsed to 0.5%, and in February 2025 DeepSeek-R1 briefly matched the top US model before closed reasoning models pulled ahead again 1. The frontier still belongs to closed labs. But the frontier is not what most workloads need.

Here is where the data turns. Mozilla and SlashData surveyed more than 1,400 developers and found that 79% of those adding AI functionality use open models, against 71% for closed 1. Adoption is not the problem. Production is. Only 51% of open-model teams reach production, compared to 63% for closed models 1. The gap is not about model quality. The top challenges developers name are infrastructure costs, security and compliance concerns, deployment complexity, and ongoing maintenance 1.

Scale does not fix it, which is the most telling number in the report. For closed models, production rates climb from 54% to 73% as companies grow larger. For open models, they barely move: 53% to 57% 1. Enterprises can buy their way through closed-model deployment. Open deployment waits on tooling that nobody has finished building.

This reframes the competitive landscape for builders and investors. If the model itself is becoming a commodity input, the companies that win the next phase are the ones solving the problems Mozilla's survey identifies: who makes open models safe to deploy, cheap to maintain, and simple to evaluate. Value is moving up the stack, to the harness around the model, not the model itself.

Mozilla's CTO, Raffi Krikorian, frames this as a familiar story. The report's opening letter draws a line to the browser wars, when one company tried to own the front door to the web and an open community stopped it 1. The argument is that open AI is running the same play.

Whether that analogy holds is a separate question. The data makes one thing clear. The question is no longer whether open models are good enough. Mozilla's numbers say they are, for most work, at a fraction of yesterday's cost. The question is who builds the infrastructure that turns capability into production.

That race is just starting.