DeepSeek has slashed prices on its artificial intelligence models, including its latest V4 which now costs 97 per cent less than OpenAI products, potentially triggering a price war in the highly competitive AI market.
DeepSeek said on Sunday that it would reduce prices for “input cache hits” – where previously processed context was reused – for application programming interface (API) users to one-tenth of the original level, bringing the minimum input cost down to about US$0.14 per million tokens.
DeepSeek said the cuts were effective immediately and would be permanent.
To promote the new flagship model released on Friday, the company offered an additional 75 per cent discount on the V4-Pro model through May 5, it said in a separate announcement on Saturday. The V4 family currently consists of the Pro and Flash variants, with the V4-Pro positioned as the company’s most advanced offering to date.
As a result, DeepSeek-V4-Pro is currently as cheap as US$0.0036 per million input tokens, a fraction of the cost of its American rivals. In comparison, OpenAI’s GPT-5.5 charges US$0.5 per million cached input tokens. Given that a conversation consists of input and output, and that input is typically three times longer than output, the cost per conversation on GPT-5.5 is 32 times that of DeepSeek-V4.
DeepSeek’s aggressive pricing strategy reflects heightened competition in China’s foundational model market, following recent releases from high-profile start-ups – including Kimi K2.6 and Zhipu GLM-5.1 – each of which raised prices on their latest flagship versions. DeepSeek V4 is one of the few models to break the trend by significantly lowering its prices, triggering speculation that the move might ignite further price competition in the market.
OpenRouter, a US-based model aggregation platform, reported a noticeable increase in usage after DeepSeek-V4 went live, with DeepSeek V4-Pro recording 13.6 billion tokens on April 25, nearly four times the previous day’s volume.
DeepSeek’s move was aimed at attracting more users, particularly enterprise clients, developers and agent-based users, Hu Yanping, a distinguished professor at Shanghai University of Finance and Economics, told the National Business Daily. He said the company was once again lowering industry pricing expectations amid recent cost increases across major model services.
Hu also noted that DeepSeek’s price cuts might have a relatively limited impact on top-tier models such as GPT-5.5 and Claude 4.7 Opus.
DeepSeek said its V4 models were optimised for mainstream agent tools, with users required to configure Claude Code by setting the model to “deepseek-V4-pro” to enable 1-million-token context support, while OpenCode and OpenClaw require an upgrade to use the latest model.
DeepSeek’s pricing overhaul comes as part of broader technical upgrades and deeper integration with Huawei Technologies’ Ascend ecosystem, which is expected to further enhance cost efficiency.
DeepSeek-V4 is seen as enabling more complex agent-based applications at a lower cost, expanding the scope for scalable AI deployment. A report from Goldman Sachs noted that integration with Ascend supernodes could further strengthen its cost competitiveness and support wider adoption.
According to an evaluation by Arena.ai, DeepSeek-V4-Pro performed at a level comparable to GPT-5.4-high and Gemini-3.1-Pro in agent-based web development tasks.
DeepSeek-V4 is far more cost-efficient than leading closed-source models such as Claude, even as its performance still lags, according to Artificial Analysis. On the third-party benchmark firm’s flagship Intelligence Index tests, DeepSeek’s Flash variant cost about US$113 in inference cost terms to run the test, while Anthropic’s latest Claude Opus 4.7 model cost around around US$4,811. -- SOUTH CHINA MORNING POST
