OpenAI and Anthropic cut prices across their API lines this week, a direct response to open-weight releases out of Beijing that now sit within a few points of GPT and Claude on the standard coding and reasoning evals (SWE-bench, GPQA). Z.ai's newest release, detailed by Wired on August 17, is the one causing the pricing move: a model good enough at offensive security tasks that researchers are already testing it against penetration benchmarks, distributed as open weights that anyone can download and run on their own servers, no API call to Anthropic or OpenAI required. That distribution model is the whole story. A lab in Singapore evaluating which model to wire into a fraud-detection pipeline can now put a downloaded Z.ai checkpoint against GPT-5.1 on the same task, on hardware it already owns, without a US vendor contract or a US export license in the loop.

Anthropic's own numbers, reported by TechCrunch on August 17, show why the price cut is defensive rather than generous: annualized revenue at $65 billion, a business Ars Technica reports could value the company at $2 trillion at IPO. That revenue comes from enterprise API contracts, the kind of deployment a bank's compliance team signs once and does not rip out for a marginally cheaper competitor. But the Z.ai release changes the marginal case: a new customer choosing a model for the first time can pick free weights over a metered API, and a US price cut only defends the customers Anthropic and OpenAI already have. The number to watch is Anthropic's next quarterly enterprise-contract count, the figure that shows whether the price war held existing accounts or is masking the first quarter where new signups went to a downloaded model instead of an API key.