Leading U.S. artificial intelligence labs, including OpenAI and Anthropic, have begun releasing artificial intelligence models that are increasingly cheaper to maintain in an effort to retain customers, who are particularly interested in lower costs and many have opted for cheap Chinese alternatives. financial times.
Source image: Mariia Shalabaieva / unsplash.com
OpenAI recently announced an 80% price cut on its “fastest and most affordable model,” the GPT-5.6 Luna, and Anthropic released the Claude Opus 5, a powerful model priced at half the price of its flagship Fable 5 model. The measures have helped customers of both companies cut costs by nearly a quarter since mid-July. It’s a new round of competition between the two leading U.S. developers of closed artificial intelligence models, which have previously only competed on performance; but the success of the Chinese rivals’ initial cheap products has pushed prices higher. OpenAI and Anthropic aim to go public at a valuation of at least $1 trillion, and they must prove to investors that big investments in artificial intelligence can generate profits.
The two companies switched some corporate customers to token consumption payment solutions, resulting in a significant increase in costs. Some responded by reducing their consumption of computing resources, and others began testing cheaper alternatives—DoorDash and Airbnb turned to the Chinese model. At the same time, Chinese developers have begun to actively improve the quality of their systems, and the gap with the United States is narrowing – there are rumors in the technology circles that American laboratories may lose customers even if they have technological advantages.
As a result, American manufacturers began to lower the prices of mid-range products to prices competitive with Chinese products. OpenAI lowered the price of GPT-5.6 Luna from $1 per 1 million input tokens to $0.20 and from $6 per 1 million output tokens to $1.2. Anthropic launched the Opus 5 for $5 and $25 — half the price of its flagship Fable 5 — and this week canceled a September price increase on the mid-range Sonnet 5 model.
Comparing on a price-per-coin basis alone isn’t entirely correct: more powerful models sometimes perform tasks with fewer resources, meaning more expensive systems can sometimes be cheaper. There are also settings for inference mode depth (effort): at medium settings, for example, the Anthropic Opus 5’s maximum results are comparable to those of the Moonshot Kimi K3. OpenAI GPT-5.6 Luna’s productivity at maximum settings is on par with DeepSeek V4 Flash, which is also set at maximum, but the US cost per task is almost half that.
If you find an error, select it with your mouse and press CTRL+ENTER.