OpenAI and Anthropic cut model prices as Chinese rivals gain users

The latest:

Prices for mid-tier models from the biggest US AI labs are falling sharply, with OpenAI slashing GPT-5.6 Luna by 80% and Anthropic pricing Claude Opus 5 at half its most capable model, the Financial Times reported. The newspaper said customers paying for leading US models have seen costs drop by almost a quarter since mid-July as Chinese developers including Moonshot and DeepSeek win users.

Details:

• The OpenAI cut: According to the FT, OpenAI reduced GPT-5.6 Luna from $1 to $0.20 per million input tokens, and from $6 to $1.20 per million output tokens, describing it as its fastest and most affordable model. The report did not say how long the lower pricing will remain in place.
• Anthropic’s pricing: Opus 5 launched at $5 per million input tokens and $25 per million output tokens, half the price of the company’s Fable 5 flagship, the FT said. The company also cancelled a planned price rise for Sonnet 5 that had been due in September. No revised schedule was disclosed.
• The index: Silicon Data’s token price index shows prices paid for leading US lab models down by almost a quarter since mid-July, per the FT. The report did not detail the index’s methodology or which models are weighted in it.
• Why customers moved: Rising AI bills have pushed some companies to cap usage or test alternatives, the FT reported, as Anthropic and OpenAI shift some enterprise clients from flat subscriptions to usage-based billing. The report did not quantify how many customers were moved off subscriptions.
• Named switchers: DoorDash and Airbnb have said they started using Chinese-made models to rein in costs, according to the FT. Neither company’s share of workloads now running on Chinese models was disclosed, and the report did not name the specific models they adopted.
• Benchmark comparison: Artificial Analysis found Opus 5 at medium effort matched Moonshot’s Kimi K3 at max effort on performance and cost per task, the FT said. GPT-5.6 Luna at max effort performed similarly to DeepSeek’s V4 Flash but cost just under twice as much per task.
• The caveat: Headline token prices are not a clean comparison, the FT noted, because stronger models can finish a task in fewer tokens or fewer attempts, and effort settings change both performance and final cost. No standardised cost-per-task figures were published across the full model range.
• The denial: A person close to Anthropic told the FT that pricing Opus 5 below Fable 5 reflects how the start-up’s family of models is built, with no connection to competitors. Anthropic and OpenAI both declined to comment on the record.
• An outside read: Mantas Lukauskas, AI tech lead at hosting provider Hostinger, said prices for the very best models were flat to rising, adding that the shift is a first real test of the labs’ ability to protect top-tier pricing: “The US labs have cut the middle and are defending the top.”
• The IPO backdrop: The cuts come while OpenAI and Anthropic plan initial public offerings at trillion-dollar valuations and investors look for evidence that heavy AI spending produces returns, the FT said. No listing timetable or filing date was reported for either company.
• إضافة: US labs are defending top-tier pricing while conceding the mid-tier, one practitioner said.

Background:

Chinese open-weight models can be downloaded and modified freely by developers, unlike the proprietary closed systems from OpenAI and Anthropic, which until now competed mainly on performance rather than price.

Between the lines:

The pattern described in the FT is selective: cuts land on mid-tier products that compete directly with Chinese open models, while Hostinger’s Lukauskas says top-end prices are flat or rising. Anthropic’s denial that competitors shaped Opus 5 pricing sits alongside the benchmark finding that Opus 5 matches Kimi K3 on cost per task.

What’s next

Watch whether Anthropic reinstates the shelved Sonnet 5 increase, how Silicon Data’s index moves through the quarter, and whether more large enterprises follow DoorDash and Airbnb onto Chinese models.