OpenAI CFO tells firms to avoid cheap Chinese AI models

OpenAI chief financial officer Sarah Friar told businesses at the International AI Summit in Dublin on Wednesday that adopting low-cost Chinese AI models carries cybersecurity risk, and that firms should stay on Western systems. "We think you want to be on more Western models, frankly," she told Bloomberg at the summit, held during Ireland's EU Council presidency.

The remarks arrive in the middle of a price war. Chinese open-weight systems have been gaining enterprise traction on cost and accessibility, ranking near the top of intelligence and cyber benchmarks while undercutting US frontier pricing. OpenAI cut its own token prices in half when the Chinese models hit the market — cuts ZeroHedge connects to July's chip-stock rout — and then delayed its IPO. Anthropic, the other major closed lab, has reduced token prices by roughly 50 percent in the past month and still charges more than OpenAI. Both companies now describe Chinese progress primarily as a security question.

Anthropic chief Dario Amodei has spent the autumn building that case, with a September essay placing China at the centre of an argument for slowing capability advances, tightening chip export controls and cracking down on distillation — the extraction of reasoning from closed US models. Parts of the technical record are documented: Anthropic's red-team testing of Z.ai's open-weight GLM-5.3 found exploit-generation performance close to its own restricted Claude model, with refusal layers weak enough to bypass with simple techniques and weights anyone can download. This week Anthropic announced safeguards it says cut distillation by roughly 90 to 99 percent, naming Z.ai's work on GLM-5.3 specifically.

Friar's warning added a purchasing recommendation. Chinese models are "getting very good from a cyber perspective," she conceded, but open weights mean "the folks with malicious intent have the tools." She also disclosed that OpenAI is holding back a next-generation model more capable than its GPT-6 Astra system while safety and alignment work continues — a model the company says produced a proof of the Navier-Stokes Millennium Prize problem, along with hundreds of other proofs, in recent weeks. "We hold back sometimes, so we talk about pacing the frontier," she said.

The model that claims a Millennium Prize sits in the vault pending safety review. The warning about cheaper rivals went straight to customers.