Why do closed AI models dominate despite costing six times more?

Frank Nagle and Daniel Yue, in a working paper summarized on the Linux Foundation's blog, report that "closed models account for roughly 80% of model usage and 96% of revenue, even though they cost, on average, six times more than competing open models." They write that this dominance "would be easy to explain if closed models consistently delivered vastly superior technical performance. But that simply isn't true": open models "routinely achieve 90% or more of the performance of closed models on widely used benchmarks" and are "closing the performance gap within a few months after each new closed-model release," with the catch-up cycle itself "accelerating." The dataset behind these figures comes from OpenRouter, which the authors say "captures nearly 1% of the global expenditure on LLM APIs."

The authors call the gap between usage and the economics that should drive it "a structural phenomenon with billions of dollars at stake," not a trivial inefficiency, and offer three reasons rather than one: "brand trust and perceived safety" that outruns the evidence for a premium; "information asymmetries," where engineering teams don't track the newest open releases or wrongly assume open means less safe; and "regulatory and liability considerations," including contractual assurances closed providers can offer and geopolitical wariness because "many powerful open models are coming out of China." They quantify the stakes by simulating a counterfactual where users pick "the best observable model—based solely on price and performance," estimating $20-$48 billion in annual unrealized savings for the global AI economy, with a preferred estimate of "$24.8 billion in annual unrealized value" extrapolated from "Menlo Ventures' 2025 estimate of the LLM inference market size" — about 70% of current LLM inference spending. Separately, they estimate that if open models disappeared entirely, consumers "would need to spend between $350 million to $1.23 billion more" than they do now, a smaller figure the authors read as consistent with decades of research on open-source software generating value through "cost reductions, complementarities, and innovation spillovers."

This sits closely alongside the companion OpenRouter-based findings that How much do open-weight models actually get used in practice? and that Does LLM price alone determine which model users choose?. Both papers draw on OpenRouter data and agree usage doesn't track price or even benchmarked capability cleanly. But Nagle and Yue push further: where the 100-trillion-token study frames the open/closed split as a "fundamental property of the market" driven by users weighing "capability, latency, price, and trust," Nagle and Yue treat the same pattern as a correctable market failure — naming brand trust, information gaps, and liability exposure as specific, addressable causes, and attaching a dollar figure ($20-48B/year) to what correcting it would be worth. Where the token-share study stays agnostic on whether the split is efficient, this paper argues it mostly isn't.

The $20-48 billion range rests on a counterfactual simulation of what "best observable" model choice would look like, not an observed natural experiment, and depends on extrapolating from a single external market-size estimate (Menlo Ventures') layered onto OpenRouter's roughly 1%-of-global-spend sample. The three causes offered — trust, information asymmetry, liability — are stated as explanations for the gap rather than tested against alternative hypotheses in the excerpt, so the paper establishes that a large usage/economics gap exists on this dataset more firmly than it establishes why. If the estimate holds up under scrutiny, it implies that an organization's "just use the frontier closed model" default is leaving money on the table by a wide margin, independent of whether open-weight models are actually weaker.