WSJ: Some Firms Still Shelling Out for Top-Tier AI Models

The interesting split showing up in enterprise AI right now, according to WSJ's CIO Journal, is not really about which model is best. It is about which companies still think that question is worth paying premium prices to answer. The reporting profiles firms that keep shelling out for the top tier even as much of the market moves the other way.

Reading around the piece fills in that other side of the split. Tech Startups reported that Western companies including DoorDash, Airbnb, and Siemens have been adopting Chinese-developed models such as DeepSeek, Qwen, Kimi and GLM to rein in inference bills. On one commonly cited workload comparison, Anthropic's Claude ran to $4,811 while Zhipu's GLM landed at $544, close to a nine-fold gap. Airbnb CEO Brian Chesky has publicly described Alibaba's Qwen as 'very good,' 'fast,' and 'cheap' for customer service tasks. Shopify itself is on record deploying Qwen 3 in a specific merchant-data extraction pipeline, reportedly cutting per-unit model costs by roughly seventy-five times.

At the same time, the capability-first buyers are still real, and that is the counter-current the WSJ piece leans on. Shopify's documented pattern of routing engineers to a wide mix of top-tier coding tools through a centralized LLM proxy, so nobody has to think about the token bill at their desk, sits neatly in that camp on the engineering surface even while cheaper models handle other pipelines. The company is not a pure-play on either side of the split, which is probably the more honest read.

The honest caveat is that the WSJ article sits behind a paywall, so the exact framing of who is in the 'shelling out' bucket and why has to be taken from the outlet's own headline rather than confirmed line by line. What the reporting also does not give you is whether the premium-paying firms have run the routed-versus-frontier comparison against their own workloads, or whether the spend is really a brand and risk-appetite decision dressed up as a quality one.

The interesting question for a CIO reading this is not which camp is right today. It is which lane your workloads actually belong in, and whether your architecture lets you move between them without a rebuild when the pricing or the quality gap moves again.