AI Race Turns into Price War: OpenAI and Anthropic Forced to Cut Prices Because of China

According to the Financial Times, OpenAI and Anthropic have begun releasing more cost-effective AI models to retain clients, many of whom have switched to cheaper Chinese alternatives. OpenAI announced an 80% price cut for its GPT-5.6 Luna model, and Anthropic introduced Claude Opus 5, a powerful model priced at half the cost of its flagship Fable 5. These measures have helped customers of both companies reduce costs by almost a quarter since mid-July. This marks a new phase of competition between the two leading American developers of closed AI models, who previously competed only on performance. The price war is fueled by the success of initially cheaper products from Chinese competitors. Both companies are aiming for IPOs with valuations of at least $1 trillion and must prove to investors that huge AI investments can generate profits. The companies have moved some corporate clients to token-based payment plans, sharply increasing their costs. Some clients reduced their consumption of computational resources, while others tested cheaper alternatives — DoorDash and Airbnb switched to Chinese models. Chinese developers have been improving quality, narrowing the gap with American systems, raising concerns that US labs might lose clients despite technological superiority. To respond, US players have lowered prices on mid-tier products to competitive levels. OpenAI cut GPT-5.6 Luna prices from $1 to $0.20 per million input tokens and from $6 to $1.2 per million output tokens. Anthropic released Opus 5 with price tags of $5 and $25 respectively, half the price of the flagship Fable 5, and this week canceled a September price increase for Sonnet 5. However, a simple price-per-token comparison may be misleading, as more powerful models can sometimes be more efficient, so a more expensive system may ultimately be cheaper. There are also reasoning depth settings: for example, Anthropic Opus 5 at medium settings performs comparably to Moonshot Kimi K3 at maximum, and OpenAI GPT-5.6 Luna at maximum settings works at the level of DeepSeek V4 Flash set to maximum, but the American model is almost twice as cheap per task.