The two US labs both dropped prices this week on their developer APIs, the pay-per-use plumbing that lets companies plug Claude or GPT into their own products. Anthropic and OpenAI cut per-token rates on their main model lines as competing systems out of Chinese labs closed the gap on the benchmarks that decide which model a company builds on: coding tests like SWE-bench, where a model has to fix a real GitHub issue end to end, not just answer a quiz question. When two vendors' models score close enough on those tests, the buying decision stops being about capability and starts being about price per million tokens, which is exactly the number Anthropic and OpenAI just moved. The timing lines up with Anthropic's own reported plan to go public at a $2 trillion valuation: a company selling API access at a discount is a company betting that locking in more developers now is worth more than the margin on each query today.

The number that decides who wins this is deployment count, not benchmark score: how many enterprise contracts, how many production API calls, how many paying seats a lab already has wired into customer systems before the Chinese labs' cheaper models show up in the same procurement conversation. A bank or logistics company that has already integrated Claude's API into its workflow does not re-architect that pipeline to save a fraction of a cent per token; the switching cost is engineering time, not sticker price. That is what protects Anthropic and OpenAI's current customer base even as the price gap narrows. What it does not protect is the next contract, the one still up for bid, where a Chinese lab's model at half the price becomes the default choice for any procurement lead who has not yet signed. Anthropic's IPO timeline, whenever it lands, will be read against exactly that number: new contracts signed per quarter, not the valuation headline.