(Yicai) May 25 -- DeepSeek will make permanent a 75 percent discount on its flagship V4-Pro artificial intelligence model, the Chinese startup has said, putting it at the top of the global price-to-performance ranking at a time when AI computing costs are generally rising.

The application programming interface price of the V4-Pro will stay at 2.5 Chinese cents (0.36 US cents) per one million cache-hit input tokens, a cut introduced on April 26 that was set to expire at the end of this month, Hangzhou-based DeepSeek announced on May 22. Cache-miss input tokens are priced at CNY3 (US$0.44) per million, while output tokens cost CNY6 per million.

DeepSeek released the V4 next-generation model on April 24. It features broad upgrades in inference performance, long-context processing, and agentic capabilities. The V4-Pro has achieved the best performance among open-source models on Agentic Coding benchmarks, with output quality that approaches Claude Opus 4.6 in non-thinking mode based on internal evaluations, according to the Chinese company.

DeepSeek’s move runs counter to a broader industry trend toward higher pricing. Amazon, Microsoft, and several major Chinese cloud providers have raised API prices by as much as 463 percent as compute costs continue to climb.

Prices for high-bandwidth memory have risen more than sixfold over the past six months, while a surge in inference demand driven by AI agents has pushed operating costs higher, making the early "cash-burning for market share" subsidy model of cloud providers unsustainable.

DeepSeek's price cut is not the result of subsidized discounting, but of structural cost savings achieved through a redesigned technical architecture, Securities Daily reported, citing analysts. Its proprietary sparse attention mechanism and mixture-of-experts architecture allow the V4 series to process million-token long-context tasks at just 27 percent of the compute cost of its predecessor, with key-value cache memory usage reduced to 10 percent, they pointed out.

The company has also extensively optimized its models for domestic AI chips, including Huawei Technologies' Ascend series, slashing hardware procurement costs, the analysts said, adding that engineering improvements on the inference side have also allowed fixed costs to be spread across a larger usage base.

The permanent price cut is also a deliberate effort by DeepSeek to fortify its position within the AI ecosystem. By sharply lowering the cost of API access, the company aims to attract developers and enterprise users to build on its platform, creating a self-reinforcing cycle of lower prices, higher usage, a stronger application ecosystem, and further cost reductions.

Editor: Martin Kadiev