Open source AI models are closing in on proprietary rivals like ChatGPT and Claude, but enterprises are still struggling to get them into production, according to Mozilla’s State of Open Source AI 2026 report.
The report – based on a survey of 1,494 developers conducted by SlashData in May 2026 and analysis of OpenRouter’s 100-trillion-token dataset – finds that open models now match closed systems on coding, instruction-following and general knowledge tasks.
Closed models still lead on reasoning and complex agentic workloads. On Terminal-Bench 2.1, the most demanding agentic benchmark included in the report, GPT-5.5 scores 83.4 percent against the best-performing open model’s 67.9 percent.
The Cost Case Is Getting Harder to Ignore
GPT-4-class inference now costs $0.40 per million tokens, down from $20 three years ago – a 50-fold drop Mozilla attributes largely to open weight model releases.
Closed models still cost roughly six times more per call at comparable performance levels, according to a Linux Foundation analysis cited in the report.
A Nagle-Yue study referenced by Mozilla estimates the unrealised annual savings from that gap at $24.8 billion across the industry.
Stripe is held up as a case study of what switching looks like in practice.
The payments company cut its inference costs by 73 percent after moving to open models running on vLLM, handling 50 million daily API calls on one third of the GPU fleet it previously required.
“Without investment in the infrastructure, tooling, and governance around open models, we risk locking in a system where only restrictive, closed AI can scale – and that doesn’t serve the public interest, or sovereignty over tech policy decisions,” said Raffi Krikorian, Mozilla’s Chief Technology Officer.
Production Deployment Remains the Sticking Point
Yet despite the economics, open models continue to lag in production.
Mozilla found that 79 percent of developers use open models, but only 51 percent have deployed them in production, compared to 63 percent for closed models.
The gap does not close at scale – open model production rates rise from 53 percent at small companies to just 57 percent at enterprise level, while closed model rates climb from 54 percent to 73 percent.
Mozilla is direct about the cause: this is an infrastructure problem, not a capability one.
Among developers who stopped using open models, integration challenges, maintenance requirements and operational complexity were the most commonly cited reasons – not performance.
The report notes that deployments completed with vendor partners reach production far more often than those built internally, pointing to a gap in enterprise-grade tooling and support around open models that the ecosystem has yet to fill.
Agent Governance Is the Industry’s Blind Spot
Agent governance is flagged as an emerging and under-appreciated risk.
Only 21 percent of companies report mature frameworks for governing AI agents.
The Model Context Protocol, which connects AI agents to external tools and data sources, has grown from 2 million to 97 million monthly SDK downloads in just 16 months, with more than 10,000 active servers now running in production environments.
Security researchers filed more than 30 CVEs in the protocol within its first eight weeks as Linux Foundation infrastructure.
Mozilla warns that while authentication has largely been solved at this layer, authorisation – defining exactly what an agent is permitted to do unattended – remains an unsolved problem with no portable standard in place.
China Is Pulling Ahead on Open Source Adoption
The geopolitical picture is shifting sharply.
Open source AI adoption in Greater China and East Asia stands at 89 percent, versus 70 percent in Western Europe.
Chinese-built open weight models now account for more than 60 percent of traffic among the ten most-used models on OpenRouter.
In February 2026, Alibaba’s Qwen model family overtook Meta’s Llama as the most downloaded on Hugging Face, with 942 million downloads against 476 million.
Mozilla describes this as deliberate industrial policy: China’s State Council AI-Plus directive explicitly promotes open weight releases as a strategy for driving global adoption while reducing dependence on US-built closed platforms.
Western governments are also responding. The EU has mobilised €200 billion for AI infrastructure and carved out exemptions for open source systems under the AI Act, while Canada has committed $890 million toward a sovereign AI supercomputer, with reducing foreign supply chain dependence as an explicit goal.
The infrastructure gap Mozilla identifies won’t close on its own – and with closed model vendors continuing to invest heavily in enterprise tooling, support and integration, the window for open source to compete on deployability as well as capability may not stay open indefinitely.
