The AI race is becoming a pricing fight. As corporate buyers scrutinise swelling model bills, OpenAI and Anthropic are finding that technical leadership alone may no longer justify a premium.

The pressure built as companies moved away from flat subscriptions and into usage-based billing, making every input and output token a line item. Businesses including DoorDash and Airbnb began testing Chinese-made models to rein in costs, while open systems from Moonshot and DeepSeek gained traction from Silicon Valley to Europe.¹

OpenAI then cut the price of GPT-5.6 Luna by 80%, taking input tokens from $1 to $0.20 per million and output tokens from $6 to $1.20. Anthropic followed with Claude Opus 5, priced at half the level of its more capable Fable 5, and withdrew a planned Sonnet 5 price increase. Across leading US labs, customer token prices have fallen by almost a quarter since mid-July.¹

The American labs argue, implicitly, that sticker prices tell only part of the story: a stronger model can use fewer tokens or require fewer attempts, lowering the cost of a completed task. Benchmarking cited by Artificial Analysis found Opus 5 at medium effort broadly matched Moonshot’s Kimi K3 at maximum effort on performance and cost per task. GPT-5.6 Luna performed similarly to DeepSeek V4 Flash at maximum effort, but cost just under twice as much per task.¹

That leaves a sharper dividing line. A person close to Anthropic said its lower Opus price reflected how its “family of models is built, so there’s no connection to competitors.”¹ But Hostinger AI lead Mantas Lukauskas framed the moment more bluntly: “The US labs have cut the middle and are defending the top.”¹

With both companies contemplating IPOs at trillion-dollar valuations, that defence is no side issue. It is the first real test of whether US AI leaders can preserve premium pricing where it matters most.¹