Open-source AI models now handle roughly one-third of all active AI workloads worldwide, yet capture only 4% of the global AI market revenue. That 29-percentage-point gap between usage and earnings reveals the central tension in AI's 2026 landscape: open-weight models have become capable enough for real production work, but the business models around them haven't caught up.

Mozilla published its first State of Open Source AI report in July 2026, offering the most comprehensive data-driven assessment to date of how open models actually compare to closed competitors like Claude Opus 4.8 and GPT-5.5. The report combined three types of evidence: a survey of 1,494 developers conducted in May 2026, analysis of 100 trillion tokens of actual model usage data from November 2024 through November 2025, and head-to-head benchmark comparisons spanning from January 2024 to March 2026.

Why Does the Usage-Revenue Gap Matter So Much?

The 33% usage figure comes from tracking actual token traffic on OpenRouter, a model router that shows which AI systems developers and companies are actually running in production. Yet open-weight models pull in only 4% of the $200+ billion global AI market revenue. Closed labs like OpenAI, Anthropic, and Google collect the remaining 96%, even though their models handle roughly two-thirds of the traffic.

This split exists because open models generate revenue differently than proprietary systems. Nobody pays Meta or Z.ai per token to run Llama or GLM weights. Instead, revenue flows to whoever hosts the inference, whether that's a cloud provider or a company running its own graphics processing units (GPUs). That value typically stays inside the buyer's own budget rather than showing up as a line item for the model creator.

The real fight of 2026, according to Mozilla's analysis, isn't whether open models are competitive anymore. It's who gets paid when they win a workload.

How Close Are Open Models to Proprietary Systems Now?

The capability gap has narrowed dramatically. Mozilla's Chatbot Arena tracking shows the average performance difference between the best open and best closed model shrank from roughly 8% two years ago to as tight as 0.5% by August 2024, before widening again once closed labs released new reasoning models. By March 2026, the gap had settled at 3.3 percentage points.

Open and closed models now perform at or near parity on several critical tasks:

• Coding Tasks: Open models match closed systems on instruction-following and general code generation work.
• Knowledge Work: General knowledge retrieval and question-answering show minimal performance differences.
• Enterprise Tools: Chatbots, internal document processing, and routine automation work equally well with either approach.

For the bulk of enterprise AI work, that parity means price and control start to matter more than a single percentage point of benchmark score. However, closed models still maintain a real edge in advanced reasoning, long-context retrieval, and agentic tasks, the kind of multi-step work that requires a model to plan and self-correct over dozens of turns.

What Changed Between the Report and Publication?

Mozilla's dataset already looked outdated on release day. The report measured the best open model at 67.9% on Terminal-Bench 2.1, a benchmark for autonomous coding tasks, against GPT-5.5's 83.4%. Weeks before Mozilla published, Z.ai released GLM-5.2, an MIT-licensed model with 744 billion parameters and a usable 1-million-token context window, meaning it can process roughly 1 million words at once.

GLM-5.2 scored 81.0 on Terminal-Bench 2.1, the first open-weight model to cross 80 on that test and a jump from its predecessor GLM-5.1's 63.5. That lands within four points of Claude Opus 4.8's 85.0. On SWE-bench Pro, a separate benchmark for resolving real-world software issues, GLM-5.2 scored 62.1, ahead of GPT-5.5's 58.6.

The model is available to run locally through tools like Ollama, and it ships with High and Max reasoning-effort settings so teams can trade response speed for accuracy on harder tasks. None of that erases Mozilla's core finding, but it sharpens it: if the best open model was already outrunning the data collected to describe it, the 3.3-point gap Mozilla measured in March probably understates where things stood by the time anyone read the report.

How to Evaluate Open-Source AI for Your Organization

• Benchmark Your Specific Use Case: Don't rely on generic benchmark scores. Test open and closed models on your actual workloads, whether that's customer support automation, code generation, or document analysis, to see which performs acceptably for your needs.
• Calculate Total Cost of Ownership: Factor in inference costs, hosting infrastructure, fine-tuning expenses, and internal support time. Open models often cost significantly less to run at scale, but require more technical expertise to deploy and maintain.
• Assess Your Control and Privacy Requirements: Open models can run on your own servers, keeping data private and giving you full control over updates and customization. Proprietary APIs offer convenience but require sending data to external providers.
• Monitor the Capability Frontier: The gap between open and closed models is closing rapidly. A model that underperforms today might be sufficient in six months, so revisit your assumptions quarterly.

Where Is Open-Source AI Growing Fastest?

Chinese open-weight models have surged dramatically. According to Mozilla's OpenRouter analysis, Chinese open models grew from under 2% of weekly traffic in late 2024 to 45% by April 2026, a shift large enough that a separate report found Chinese AI models had overtaken US rivals in enterprise deployment share.

Regional adoption patterns show stark differences. Open-source AI utilization runs at 89% in Greater China and East Asia, against 70% in Western Europe. On Hugging Face, a popular repository for open models, Alibaba's Qwen models have logged 942 million downloads and Meta's Llama family 476 million, a rough proxy for how far open weights have spread into ordinary developer workflows.

What Has Happened to AI Pricing?

Inference costs have collapsed. GPT-4-class inference fell from about $20 to roughly $0.40 per million tokens over 36 months, a 50-fold drop that Mozilla ties directly to open-weight competition pushing prices down across the board.

That price compression affects both open and closed models. Even proprietary systems cost far less to run today than they did three years ago, a trend that accelerated as open alternatives proved capable enough to handle real workloads. The economics of AI have shifted from "pay per token to a closed lab" to "choose the model that fits your budget and control requirements, then pay for the infrastructure to run it."

Mozilla's report lands at a moment when that choice has become genuinely difficult to make. Open models are no longer the scrappy underdog option for cost-conscious developers. They're a legitimate production choice for most enterprise AI work, and the gap keeps narrowing on the remaining specialized tasks where closed models still lead.

" }