Chinese startup DeepSeek has announced that the planned 75% promotional discount on its flagship V4 Pro model will not be temporary. It will remain in place permanently. Prices for developers have thus fallen to a quarter of their original level, and the Chinese company is now undercutting virtually all the largest Western artificial intelligence providers.
From a temporary promotion to permanent pricing
When DeepSeek launched the V4 Pro model in April 2026, it openly admitted that prices were temporarily higher. The reason? Limited availability of computing power. At the time, the company explicitly stated that the Pro version would cost up to twelve times more than the less powerful Flash version due to a shortage of cutting-edge chips. It also indicated that prices would fall significantly once Huawei released larger quantities of its Ascend 950 compute nodes. A 75% discount followed. It was originally intended to remain valid only until the end of May. However, DeepSeek made it permanent on May 23.
The new prices range from 0.025 to 6 yuan per million tokens, or roughly $0.0035 to $0.83. The original rates ranged from 0.1 to 24 yuan.
OpenAI's GPT-5 charges $2.50 per million input tokens and $10 per million output tokens. Anthropic charges $5 for input and $25 for output for its Claude Opus 4.7 model. Google Gemini 3.5 Flash, its more cost-effective option, costs $0.15 for input and $0.60 for output.
At its permanent prices, DeepSeek V4 Pro falls below all these figures. The greatest difference is evident among frontier models designed for demanding enterprise deployments, where token consumption scales fastest. Salesforce estimates that it will spend more than $300 million this year on Anthropic tokens alone. At DeepSeek's prices, the same volume would cost only a fraction of that amount.
Huawei in the background, Washington paying attention
DeepSeek did not say whether the permanently lower prices are directly related to the growing availability of Huawei Ascend 950 chips. What is certain, however, is that V4 was built around these chips from the outset, and the company makes no secret of their impact on the model's performance.
Huawei is seeing growing demand precisely because U.S. export restrictions prevent Nvidia from selling its most advanced chips on the Chinese market. However, the same restrictions also affect equipment used to manufacture the chips themselves, making it difficult for Huawei to significantly increase production of Ascend chips. It is not yet clear how quickly the situation will change.
From the end of 2025 to April 2026, Anthropic increased its annual revenue from $9 billion to $30 billion, largely thanks to enterprise adoption of the Claude Code tool. DeepSeek is now targeting this very segment.
According to analysts, companies will begin sending less demanding tasks to DeepSeek while continuing to use Claude or GPT-5 for the most sensitive and complex queries. This would mean that the total token volume at Anthropic or OpenAI would not decline, but the average price per token would. And that is precisely the metric supporting their valuations.
Meanwhile, OpenAI is adding consumer features, such as personal finance tools, and is trying to build a business model that is less dependent on API revenue. With a valuation exceeding $850 billion, it cannot afford to wait and see how the price war plays out.
An advantage that comes at a price
The decision to adopt DeepSeek V4 Pro is not a simple equation for chief technology officers. The model offers a one-million-token context window at prices unmatched among frontier models. This is attractive for processing extensive documents, conducting legal research, or analyzing code—precisely the areas where input costs quickly add up.
On the other hand, there is geopolitical sensitivity. Routing corporate data through a Chinese provider raises questions about regulatory compliance, data protection, and operational reliability. In the government sector or regulated industries such as finance and healthcare, this is a factor that significantly complicates the decision.
Anthropic has also publicly accused DeepSeek of so-called "distillation attacks," alleging that DeepSeek trained its models on Claude's outputs to improve its own performance. DeepSeek has not responded to the allegations in detail. If confirmed, DeepSeek would be building part of its price advantage on research by others that it never paid for.
The pricing is intended to attract the greatest possible volume of usage regardless of the margin on each token. In its statement, the company described V4 as ushering in the "era of the affordable one-million-token context window." It is a direct message to developers who have so far been conserving tokens.
Whether this strategy will also succeed outside the Chinese market depends on how quickly competitors can respond. Google has repeatedly reduced Gemini's prices to counter open models. However, neither OpenAI nor Anthropic has yet taken a similarly dramatic step.
Sources: bloomberg.com and thenextweb.com
