OpenAI (OPAI.PVT) CFO Sarah Friar argued Tuesday that the company’s lightweight Luna model is far less expensive than competing open-weight Chinese models, while still offering strong overall performance.
Open-weight models, or models that users can customize and run on their own hardware, are seen as a cost-saving measure for businesses that want to spend less on AI without having to cut back on using the technology. And China dominates the market.
OpenRouter, which tracks real-world usage of models accessed through its model-picking platform, shows seven of the top 10 models by token usage over the last month were open-weight models from China, with DeepSeek and Z.AI taking the first, second, and third spots.
OpenAI’s Luna was the fourth-most-used model, up 115%.
“A model like Luna is actually about one-fifth, one-sixth the cost of, say, a Chinese open-source model,” Friar told Deirdre Bosa.
“That is one of the myths of 2026 that I think we have to myth-bust,” she added.
Friar explained that OpenAI saw a 10x jump in Luna usage after dropping its price by some 80% just two weeks after launching the model.
In an interview with Yahoo Finance, Santa Clara University Leavey School of Business professor Ram Bala estimated that, in the future, open-weight models will account for about 80% to 90% of total token usage, but closed models will make up the vast majority of dollars spent.
Open-weight model makers generate revenue by hosting their models on their own hardware and having customers pay to access them. Users who have access to powerful AI systems can, however, download the models for free and run them on their own.
To better compete with open models and gain an edge on their closed-weight competitors, frontier labs like OpenAI and its rivals Anthropic (ANTH.PVT), Google (GOOG, GOOGL), and Meta (META) have kicked off a pricing war, with each company working to improve its overall model and computing efficiency and reduce the cost to use its services for customers.
Email Daniel Howley at dhowley@yahoofinance.com. Follow him on X at @DanielHowley or reach out on Signal or Telegram.
