OpenRouter's Anthropic–OpenAI Split Nears Parity

OpenRouter says Anthropic collected about 75% of business spending on Anthropic and OpenAI models in January 2026; by September, its share was near half. Developers can keep a coding interface and change the model behind it. The spending figures cannot tell whether those developers drove the shift.

OpenRouter's spending split moved inside one marketplace

OpenRouter's business-spending figures track only two providers on one service. They do not measure Anthropic's share of all enterprise AI purchases.

OpenRouter also reported that its users spent more on OpenAI models than Anthropic models in the week of September 7, the first such week since February 26, 2024. The weekly result covers a different interval and user group from the monthly business-spending comparison. Neither measure follows a named customer's task from one model to another. New customers, changing prices, and changing usage could all move the dollar split.

Ramp captures different purchases. In March, Axios reported that Anthropic took about 73% of spending among Ramp customers buying AI tools for the first time, up from a roughly even split with OpenAI in January. Ramp's July AI index put Anthropic's share at 43.5%, widening its lead over OpenAI. Neither Ramp's sample nor OpenRouter's two-provider marketplace measures an enterprise-wide switch.

Traffic tells OpenRouter what buyers select, not why

In May, coverage recorded OpenRouter's reported throughput at 25 trillion tokens a week across more than 400 models, up from 5 trillion six months earlier. At that scale, OpenRouter can observe demand moving among suppliers inside its marketplace. A directory can show that a model exists; a purchasing interface can show whether buyers keep paying for it when alternatives sit beside it.

Prices and supplier costs can change after a buyer settles on a model. On July 1, The Information reported that OpenAI engineers had told colleagues they had found a way to more than halve inference costs; the report did not establish that OpenAI deployed the change. Reuters reported on August 21 that OpenAI cut GPT-5.6 Sol's API and credit prices by more than 20% for three months, to $4 per million input tokens and $20 per million output tokens. A lower listed price gives a buyer reason to check a choice again, but does not establish the cost of a completed task.

OpenRouter's reported volume does not reveal whether its service automatically selects models for customers, how any such selection weighs price against latency or reliability, or whether a customer saved money on a completed task. Buyers need those comparisons to turn a spending record into buyer power. Token counts alone cannot distinguish a cheap useful answer from a cheap retry.

Developers are separating the work surface from the model

Some developers reportedly run Claude Code through proxies and services including OpenRouter to use cheaper non-Anthropic models, among them OpenAI's GPT-5.6 Sol. They keep a familiar coding harness while changing its supplier. The coverage does not count those developers or show that OpenRouter picked their models.

Cursor made a different separation for Composer 2. Moonshot says its Kimi K2.5 provides the offering's foundation and that Cursor accesses the model through Fireworks AI. Cursor's integration attaches a named model to a named product; it does not make Kimi interchangeable with every other model. Product fit still matters even when the application, access provider, and model maker are different companies.

An unknown model can win attention before it wins a contract

Z.ai tested GLM-5.3-Flash anonymously as Ox Alpha on OpenCode and OpenRouter before disclosing its identity and releasing the weights. Ox Alpha offered free access and advertised a one-million-token multimodal context window. It topped OpenRouter's leaderboard while users knew it as a model from an unknown lab.

The anonymous trial put Ox Alpha in front of developers before Z.ai revealed who made it. A leaderboard lead at zero price says nothing about paid conversion or quality on a particular customer's task. The trial does show that developers could assess an unfamiliar model on the marketplace before choosing its vendor.

Stripe wants the marketplace; Meta is building its own router

In August, Stripe agreed to acquire OpenRouter for a reported $7.5 billion. Stripe's payments business gives the proposed deal an economic logic: token purchases vary by workload and supplier, and someone must measure and settle the bill. The agreement does not establish that the acquisition has closed, that Stripe has combined payments with task-level performance measurement, or that OpenRouter can dictate model prices.

Meta's internal AI incubator, meanwhile, is developing Switchboard to send some tasks to lower-cost models. Meta can try to route its own workloads; if Stripe completes its purchase, a payments company would own a marketplace handling many customers' model purchases. Neither report shows a buyer comparing the cost and result of sending the same task to different models.

For a developer keeping Claude Code but changing the model behind it, the open question is whether the same coding task costs less without a worse result; neither OpenRouter's spending figures nor the reported proxy use answers it. The fall from 75% to near half tells which provider collected the dollars, not whether a buyer got more for them.