The AI Price War: OpenAI and Anthropic Cut Prices Under Pressure from China

The largest American AI labs have entered a price war. OpenAI and Anthropic have begun rolling out cheaper models in an effort to retain cost-conscious customers who are increasingly switching to inexpensive alternatives from Chinese developers.

The reason is simple: mounting AI bills are pushing companies to rein in their use of these services and hunt for cheaper options. Against this backdrop, Chinese developers, including Moonshot and DeepSeek, are finding users from Silicon Valley to Europe.

Who cut prices, and by how much

OpenAI slashed the price of GPT-5.6 Luna — its “fastest and most affordable model” — by 80%: from $1 to $0.20 per million input tokens, and from $6 to $1.20 per million output tokens. Anthropic introduced Claude Opus 5 at $5 and $25 per million input and output tokens, respectively — half the price of its flagship Fable 5. This week the company also canceled a price hike for Sonnet 5 that had been planned for September.

According to the Silicon Data index, the prices customers pay for the models of leading American labs have fallen by nearly a quarter since mid-July. It's a striking reversal for companies whose proprietary “closed” models had previously competed mainly on performance rather than cost. Much of the downward pressure on prices comes from open Chinese models, which can be freely downloaded and fine-tuned.

A trillion-dollar IPO and the search for returns

The price cuts coincide with plans by OpenAI and Anthropic to go public at valuations of around a trillion dollars each — even as investors wait for proof that the colossal investments in AI can turn a profit. At the same time, both companies are shifting some of their corporate clients from fixed subscriptions to paying for the compute resources they consume. Some firms are responding by capping their AI usage or testing cheaper solutions: DoorDash and Airbnb, for example, have already said they are turning to Chinese models to save money.

Comparing models by the “list” price of tokens, however, is misleading. A more capable model often solves a task using fewer tokens or in fewer attempts, and “effort” settings change the amount of compute expended. According to Artificial Analysis, Opus 5 at “medium” effort delivered a result and per-task cost close to Moonshot's Kimi K3 at “maximum,” while GPT-5.6 Luna came in at twice the price of DeepSeek V4 Flash for comparable quality.

Meanwhile, prices for the most powerful models remain stable or are rising. As Mantas Lukauskas of Hostinger put it, the American labs have “cut the middle and are defending the top” — and the current changes will be the first real test of whether they can hold the line on the cost of their flagships.

Source: Ars Technica