For developers who track every fraction of a yuan per million tokens, the notice landed like a rate cut from a central bank. DeepSeek announced this week that the promotional price on its V4-Pro model API will not expire. When the limited-time discount ends on May 31, the price will simply become the permanent list price — one-quarter of the original.
The new rates are striking in their precision. Input tokens that hit the cache cost 0.025 yuan per million — two and a half cents per million tokens. Uncached input costs 3 yuan per million, and output 6 yuan per million. At those levels, DeepSeek has set what industry observers describe as the new global floor for frontier-model pricing, undercutting American rivals by a wide margin on comparable quality.
A discount becomes the price
The mechanics of the announcement matter as much as the numbers. DeepSeek had been running a 75% discount on V4-Pro since the model’s launch in late April, a promotional rate that was scheduled to revert to full price in June. Instead, the company said the discount is permanent — converting what looked like a temporary campaign into a structural pricing position.
The rate card covers the V4-Pro model, the largest in DeepSeek’s current line, which the company released open-weighted in April with a 1-million-token context window. The company’s consumer apps remain free, so the change hits only API customers: developers and enterprises building products on the model. For those users — especially teams running heavy workloads with high cache-hit rates, such as retrieval-augmented generation systems — the effective cost drops by more than 90% against the original list price.
A pricing war with an agenda
The move lands in a market where price has become the sharpest competitive weapon. American frontier labs have been pulling prices down as they scale, but DeepSeek’s cost structure, built on aggressive optimization and an open-weight distribution strategy, allows it to price where Western rivals say they cannot follow. Analysts describe the permanent cut as a strategic statement: DeepSeek is choosing ecosystem share over near-term revenue, betting that cheap access to a strong model will build the developer base that matters in the long run.
The timing is also notable. Bloomberg reported this week that DeepSeek is advancing a financing round of about 70 billion yuan, roughly $10 billion — potentially the largest single fundraising by a Chinese AI company — with founder Liang Wenhong reportedly planning to invest about 20 billion yuan of his own money. The price cut and the fundraising go together: the company is spending to win scale, and investors are funding the campaign.
What it means for developers
For application builders, the effect is immediate. Teams that had budgeted for higher inference costs can now plan around a cheaper floor, and the gap between DeepSeek’s price and the cheapest comparable U.S. models — several times higher per token — makes it the default option for cost-sensitive workloads. The company’s API is compatible with the OpenAI format, which lowers the switching cost for developers already integrated with Western providers.
The longer-term implication is a two-speed market. At the high end, American labs continue to sell premium models at premium prices to enterprises that value safety, support and compliance. At the low end, DeepSeek is defining the commodity price of intelligence — and forcing every competitor to decide whether they can match it. The answer to that question will determine where the industry’s margins land over the next two years.
The discount’s history shows how far the company has come in weeks. When DeepSeek released its V4 line in late April, open-weighted and free to download, it paired the launch with an API pricing adjustment that cut cache-hit input prices to a tenth of their original level, then added a limited-time 75% discount on V4-Pro. The promotion was scheduled to expire at the end of May. Making it permanent removes the uncertainty that developers face when building on a promotional rate — the risk that the economics of a product change when the campaign ends.
The positioning is deliberate. DeepSeek has built its reputation on open-weight models and radically low prices, and the permanent cut extends that identity from research into commercial infrastructure. The company’s stated ambition is artificial general intelligence, and it has said it will keep releasing open models while it pursues that goal. Cheap API access is the flywheel: more developers, more usage, more data on how the model is used, better products, lower costs — a cycle that Western labs, with their focus on enterprise sales and premium pricing, have approached differently.
The competitive response will shape the market. American labs have argued that their prices reflect safety investment, enterprise support and reliability — costs DeepSeek, with its leaner operation, does not carry in the same way. If customers conclude that quality differences no longer justify the price gap, the frontier of the AI market could shift from who has the best model to who has the lowest cost of serving it. That is the question DeepSeek’s permanent cut puts to the entire industry, and the answer will be written in next quarter’s market-share numbers.
DeepSeek has turned a promotional discount into a permanent price position, setting a new global low for frontier-model API pricing. The move is a bid for developer share financed by record fundraising — cheap tokens today, ecosystem dominance tomorrow. For the rest of the industry, the message is uncomfortable: the cheapest frontier model is no longer a promotion; it is the standing price.
