This week Anthropic extended free access to Claude Fable 5 for the second time in seven days, three days after OpenAI shipped GPT-5.6 and its flagship Sol model with claims of state-of-the-art results in coding, science and cybersecurity. It is the kind of exchange the industry treats as the main event: two labs, two flagships, competing benchmark claims, a promotional giveaway as a defensive move. But a set of numbers published the same week suggests the main event may be happening somewhere else entirely, and that the people fighting over the frontier are fighting over a shrinking share of the actual work.
Chinese open-weight models accounted for 41 percent of downloads on Hugging Face this spring, overtaking US models. On OpenRouter, which routes requests across providers and publishes usage data, the six most popular models are now all open models from Chinese firms including Tencent, Xiaomi, DeepSeek, MiniMax and Z.ai. Anthropic's Claude Opus 4.7 sits seventh. Data from Vercel shows open-weight models handled nearly a third of AI requests on the platform in June, absorbing the volume-heavy infrastructure work while closed models operate as a premium layer above it.
The caveat matters and is worth stating plainly: these platforms capture only a slice of the market. They exclude sessions hosted directly by the major labs, which almost certainly account for the bulk of OpenAI's and Anthropic's usage. Nobody should read an OpenRouter leaderboard as a verdict on who is winning. But the direction is consistent across three independent sources, and it lines up with what enterprises say they are doing.
What they are doing is defecting on price. DoorDash, Airbnb and Siemens have all adopted Chinese AI tools, according to the Financial Times. DoorDash cofounder Andy Fang said the company is saving substantially by routing "lower-level work" to a model from Moonshot AI. The San Francisco startup Lindy has dropped Anthropic's models entirely in favour of DeepSeek's V4. "Enterprises are starting to realize, hey, we don't need the best model, we can use the faster, cheaper models," Featherless AI chief executive Eugene Cheah told the FT, calling Chinese models "the elephant in the room."
That sentence is the whole story compressed. The frontier labs have spent three years arguing, correctly, that their models are the most capable. The argument that is now beating them is that most production work does not need the most capable model. Sam Bresnick of Georgetown's Center for Security and Emerging Technology put the enterprise logic bluntly: why pay a premium for Anthropic or OpenAI when Chinese models are workable for a lot of the workloads you actually run?
Cost is the trigger, but it is not the only pull. Open weights can be fine-tuned, self-hosted, and inspected. Hugging Face chief executive Clem Delangue argues that companies increasingly want to own their models rather than rent them, and that the industry is heading away from a "one model to rule them all" future toward companies running many models customised for specific uses. A new repository is created on Hugging Face every seven seconds; the platform hosts roughly three million public models, and half the Fortune 500 use it to deploy private or open models. "You don't want to outsource your core capabilities to a black box API," Delangue said.
There is an obvious tension here that neither camp has fully confronted. The open models eating the volume layer are trained, in part, in the shadow of the frontier models they are displacing, and the frontier labs are the ones absorbing the enormous cost of pushing capability forward. If the economics of that push depend on capturing the production workload, and the production workload keeps migrating to whatever is cheap and good enough, then the business model funding the frontier is being undercut by its own downstream. Anthropic has already accused Chinese open source models of distilling Claude by sending it millions of prompts.
None of this makes the Fable-versus-Sol contest irrelevant. Frontier capability still sets the ceiling, and specialised high-value tasks will keep reaching for it. But there is something telling about a week in which the most-watched fight was a free-access promotion between two labs, while the models quietly running most of the work belonged to neither of them.