DeepSeek makes 75% price cut permanent on V4-Pro as AI price war enters new phase

DeepSeek will permanently keep the 75% discount on its flagship V4-Pro model that had been set to expire May 31, the company said Saturday. V4-Pro API output tokens now cost $0.87 per million, down from $3.48 at launch one month ago.

DeepSeek announced Saturday that the 75% discount on its flagship V4-Pro model is now permanent, dropping API costs to as low as $0.0035 per million tokens and intensifying a price war with OpenAI, Anthropic, and Google.

The Chinese AI startup had originally set the promotion to expire on May 31, according to its pricing page. Instead, the company said in a statement that V4-Pro API pricing would stay at a quarter of launch rates indefinitely, with costs ranging from 0.025 to 6 yuan ($0.003625 to $0.87) per million tokens depending on usage type.

Claude charges 29 times more for the same output token

With the new rate structure, the V4-Pro output tokens cost $0.87 per million tokens. Claude Opus 4.7 costs $25 per million tokens. The cost of GPT-5.5 is $30.

According to CloudZero, Google’s Gemini 2.5 Pro costs $12. Other lower-cost Western LLMs, such as Claude Sonnet ($15) and GPT-4.1 ($8), are 9 times to 17 times more

An application generating 100 million output tokens per month would cost roughly $87 on V4-Pro at the new rate, versus $2,500 on Opus 4.7 or $3,000 on GPT-5.5, per MindStudio.

As Cryptopolitan earlier reported when V4 launched last month, DeepSeek said the Pro version would initially cost up to 12 times more than the lighter Flash version due to “constraints in high-end compute capacity.”

The company added that pricing would fall once Huawei Ascend 950 supernodes shipped in larger quantities in the second half of 2026. The permanent cut arrived months before that target.

DeepSeek did not say whether Huawei chips made the cut possible

DeepSeek did not disclose whether the permanent price cut was due to the increased supply of Huawei’s Ascend 950 chips.

Huawei’s AI chip sales have benefited from US export controls that prevent Nvidia from selling its most advanced semiconductors in China. Separate restrictions on chipmaking equipment have limited Huawei’s ability to scale Ascend production.

Huawei is targeting 750,000 AI chip shipments in 2026, per industry estimates, though its total production amounts to roughly 3% to 5% of Nvidia’s combined computing power.

Four Chinese chip companies (Huawei Ascend, Cambricon, Hygon Information, and Moore Threads) confirmed their hardware worked with V4-Pro from launch day.

The V4 launch already triggered a selloff in Chinese AI stocks

In early April when V4 was introduced, the share prices of some Chinese artificial intelligence firms fell on the Hong Kong stock exchange. Zhipu AI declined by about 8%-9% while MiniMax saw a fall of 7%-8%.

According to CNBC, Manycore Tech fell by 9%. Stocks of chipmakers rose contrary to the trend.

DeepSeek’s pricing strategy has also drawn friction with Western competitors. Anthropic has previously accused the company of “distillation attacks” that improperly learn from Claude’s models, per Engadget.

The White House accused China last month of stealing US AI labs’ intellectual property. The Chinese Embassy in Washington rejected those claims.

DeepSeek is simultaneously in talks to raise outside funding for the first time, with Tencent and Alibaba among potential investors at a valuation above $20 billion, per Cryptopolitan.

“Crypto bots and DeFi agents finally have pricing that works”

Agents that execute multi-step tasks autonomously can burn through millions of tokens per session.

The daily cost for running an agent such as Claude Opus at a rate of $25 per million output tokens is hundreds of dollars. DeepSeek V4-Pro at $0.87 is under $40.

Crypto trading bots, DeFi watch agents, code reviewing processes, and market analysis applications all rely on the inference costs remaining below a threshold where the value of the output is greater than the compute cost.

The announcement by DeepSeek about its permanent pricing enables a category of agent applications to be feasible on its platform that are not feasible on Western platforms at the present rate.

OpenAI has been lowering its prices throughout the last year. Anthropic released Claude Haiku 4.5 for $1/$5 per million tokens. Google has rolled out Gemini Flash for a $0.30 input price. However, none of them could compete with DeepSeek’s pricing model.

The V4-Pro is not a budget model. It has 1.6 trillion parameters, a 1 million token context window, and benchmarks that sit close to GPT-5.4 on math and reasoning tasks, per MindStudio analysis.

Don’t just read crypto news. Understand it. Subscribe to our newsletter. It's free.