OpenAI and Anthropic are cutting prices on some AI models as lower-cost Chinese alternatives gain traction with cost-sensitive customers, according to Ars Technica. The report frames the moves as a price war among leading US AI labs that had previously competed primarily on model performance rather than cost. The clearest price move comes from OpenAI. Ars reports that OpenAI recently said it was reducing prices for GPT-5.6 Luna, which it described as its “fastest and most affordable model,” by 80%. The model’s price fell from $1 to $0.20 per million input tokens and from $6 to $1.20 per million output tokens. Anthropic has also moved downmarket on price. According to Ars, the company launched Claude Opus 5 at $5 per million input tokens and $25 per million output tokens, describing it as offering frontier-level intelligence at half the price of Fable 5, its most capable model. Ars also reports that Anthropic this week called off a planned price increase for Sonnet 5 that had been due to take effect in September. The cuts appear to be showing up in customer costs. Ars cites Silicon Data’s token price index, which indicates that prices customers pay for models from leading US labs have fallen by almost a quarter since mid-July. Tokens are the billing units used by many language-model providers: input tokens measure data sent into a model, while output tokens measure the text or other response generated by the system. The pressure is coming from both sides of the market. On the supply side, Ars reports that Chinese developers including Moonshot and DeepSeek are winning usage from Silicon Valley to Europe with cheaper models. The report says increasingly capable open Chinese models, which developers can download and modify, have narrowed the performance gap with leading US systems and contributed to pricing pressure. On the demand side, corporate AI buyers are facing more direct exposure to usage costs. Ars reports that Anthropic and OpenAI have shifted some enterprise customers away from flat subscriptions and toward usage-based billing, where charges are tied to the compute resources consumed. Some businesses have responded by limiting internal AI usage or testing lower-cost alternatives. Ars names DoorDash and Airbnb as companies that have said they started using Chinese-made models to control AI bills. The evidence in the provided material does not quantify how much those companies are using Chinese models, or whether those deployments are replacing US models or supplementing them. The market context matters for the US labs’ financing story. Ars reports that OpenAI and Anthropic are preparing for possible initial public offerings at trillion-dollar valuations, while investors look for proof that heavy AI infrastructure spending can produce durable returns. Lower model prices could help retain customers, but they also put more pressure on the economics of serving those customers. The report also cautions that headline token prices are not a complete model-cost comparison. More capable models may use fewer tokens or require fewer steps for a given task, while different model versions and “effort” settings can change the effective bill. For buyers, the relevant metric is not just price per token, but the total cost of completing a workflow at acceptable quality. Who benefits: Enterprise AI buyers benefit from lower listed prices and more model options. Chinese model developers including Moonshot and DeepSeek may benefit if cost pressure keeps pushing companies to test alternatives. Who's exposed: US labs with high compute spending and ambitious valuation targets are exposed if price cuts become necessary to retain usage. Customers relying on usage-based billing are also exposed to unpredictable AI bills unless they manage consumption carefully.