OpenAI and Anthropic dropped API prices this week as Chinese model providers took paying customers, and the pressure traces to one number: DeepSeek and Alibaba's Qwen have been quoting roughly a tenth of GPT and Claude's per-token rate since the spring, and enterprise buyers comparing outputs on coding and support tasks have started to notice the gap in quality is smaller than the gap in price. A procurement team switching a customer-support pipeline from GPT to a Chinese model doesn't need the Chinese model to be better. It needs it to be close enough, and every quarter it stays close enough, OpenAI and Anthropic lose the margin they were counting on to fund the next training run.

The number that makes this bite is the cluster cost, not the API price. Training GPT-5-class or Claude-class models runs on tens of thousands of Nvidia B200 GPUs, hardware that costs on the order of $2-3 billion per large training cluster once power and cooling are counted. That capital gets repaid out of API margin. Cut the price to match DeepSeek and the repayment window stretches; hold the price and lose the accounts price-sensitive buyers control. OpenAI's Denise Dresser departure this week is a boardroom footnote next to that math, but Anthropic's reported $2 trillion IPO valuation now has to survive a market where its own product line is discounting against itself. The quarter that tests this is the one where Anthropic's public filings first have to show whether API revenue per dollar of compute is rising or falling, and that filing is coming, not later than the IPO roadshow itself.