OpenAI climbs to about half of shared customers' AI spending after cutting its smallest model's price 80%

Spending share is counted in dollars, so a price cut works against the company making it. At 80% off, a dollar buys five times as many Luna tokens as before ^[25]. For OpenAI's slice of the dollars to grow after that cut, the volume moving its way had to rise faster than its prices fell, or Anthropic's dollars had to shrink ^[2][5]. Among the shared customers, OpenAI's spending per dollar going to Anthropic went from about 33 cents to about a dollar, a threefold rise ^[26]. OpenRouter reported shares, so the figure is relative and does not show whether either company's total grew ^[2].

Price is the explanation the Journal leads with. As summarized by Trending Topics, it said the price war was fueled mainly by the GPT-5.6 lineup ^[4]. Retool says it spends about 20% less on OpenAI models than on comparable Anthropic ones ^[11]. "The GPT-5.6 opened up this new lane for OpenAI in a way that the Anthropic family of models doesn't really have," said Peter Walker, head of insights at OpenRouter ^[1].

Supply is a second explanation. Demand for Claude Code was strong enough in the spring to cause frequent outages and a computing crunch ^[12], and some Claude users balked at guardrails on the Fable 5 model released in early June ^[13]. A customer who cannot get capacity spends the money elsewhere, whatever the price list says.

Hedging is a third. When one vendor starts the year with three-quarters of the spend ^[3], any customer that diversifies moves dollars to the other one. David Zhu, chief executive of Reevo, said his company is shifting toward OpenAI for cost and to be less reliant on any one model maker ^[14]. "The honeymoon phase of being tied to one model maker is gone," Zhu said ^[15].

We think price was the main cause inside this sample, and that the sample overstates how much price decides elsewhere. OpenRouter, which Stripe bought this summer for $7.5 billion ^[20], routes requests to providers by price and performance ^[9], so its users have already chosen a tool built for switching. Ramp's data skews toward high-growth, tech-forward companies ^[8]. At the top end, Michael Szklarski of the videogaming startup ReadyM said his company prefers Anthropic's Fable 5.1 and hoped to talk to Anthropic engineers about bringing the cost down ^[23].

The test of that view is already running. Haiku 5.5's cut works out to about 90% ^[27], deeper than Luna's. If OpenAI holds near half of shared spending once small-model prices are level, capacity or hedging explains more of its gain than the cuts did. If it slides back toward a quarter, the cuts explain it.

The cuts came after OpenAI decided what to stop doing. In the spring it shut down side projects, including the Sora video app, to concentrate on businesses ^[21]. "We cannot miss this moment because we are distracted by side quests," Fidji Simo, then OpenAI's chief executive of applications, told staff, according to the Journal ^[22].