DeepSeek has introduced discounted off-peak pricing for developers using its AI models to build their products.
Per a pricing table on its website, API usage costs for the R1 model will drop by 75%, while the V3 model will be 50% cheaper during this period.
The Hangzhou-based AI startup is responding to surging demand from users and developers, who are rapidly adopting its technology.
DeepSeek disrupts the AI market with aggressive price cuts
DeepSeek’s affordable AI models sparked a major sell-off in global equity markets last month, as investors feared they would impact existing AI market leaders.
According to the Hangzhou-based company, from 16:30 GMT to 00:30 GMT, the cost of using its API— which enables developers to integrate its AI models into their applications and web products— would be reduced by up to 75%. The move could pressure rivals in China and abroad to lower their prices.
Although DeepSeek labels this timeframe as “off-peak” since it runs from 00:30 to 08:30 Beijing time, it coincides with daytime hours in Europe and the United States, where the company’s low-cost yet powerful AI models recently triggered a sell-off in tech stocks.
DeepSeek’s latest price discounts on Wednesday mark another bold disruption move in the AI industry in China and globally. According to sources familiar with the company, DeepSeek is now fast-tracking the launch of a successor to its R1 model, which debuted in January.
Competitors like Tencent Holdings Ltd. and Perplexity AI have begun using DeepSeek’s models, joining state entities across China, including departments in the Hong Kong government.
On Tuesday, DeepSeek reopened top-ups for credits to access its API but warned that capacity would be limited during peak daytime hours.
DeepSeek sparks AI price war with its low-cost model
Since its launch on January 20, DeepSeek’s reasoning chatbot has challenged conventional assumptions about the cost of developing advanced AI models. Praised as a potential rival to OpenAI, it delivers high performance at a fraction of the cost.
The model’s success has fueled a rally in Chinese tech and internet stocks, as investors anticipate further gains despite ongoing chip trade restrictions from Washington.
The company’s commitment to open-source development and aggressive pricing first ignited a domestic AI price war last May with the release of its V2 model, the predecessor to the AI powering its widely used assistant.
Since DeepSeek launched its AI assistant last month, OpenAI, the maker of ChatGPT, has cut prices, while Google’s Gemini has introduced discounted access tiers.
DeepSeek cuts AI costs while matching industry giants
DeepSeek developed one of the world’s most powerful AI systems while using significantly fewer resources than its competitors. Unlike leading AI companies that rely on supercomputers with over 16,000 GPUs, DeepSeek achieved similar results with just 2,000, reducing its raw computing costs to $6 million—roughly one-tenth of what Meta spent on its latest AI model.
The company achieved this cost efficiency using a mix of innovative techniques. One of its primary tactics was the “mixture of experts” technique, which divides the AI system into multiple specialized neural networks, each focusing on a specific skill, such as poetry, programming, physics, or biology.
Instead of streaming all data through a single large model, these smaller expert systems operated independently while a “generalist” network coordinated their interactivity. This greatly lowered the computational cost and avoided expensive data transfers between GPUs.
The Chinese AI start-up then further optimized its own AI training via numerical precision reduction. Rather than store numbers in the customary 16-bit memory format, the company crunched them down to 8-bit memory, halving its storage needs. While this slightly reduced the precision of individual calculations, the tradeoff had little effect on overall model performance.
The company’s final calculations were in a 32-bit format, which allowed it to keep efficiency while still having accuracy where it mattered.
DeepSeek also had the advantage of writing very efficient team software that best used the GPU. Although many AI labs hire skilled engineers, few have successfully rolled these cost-saving techniques out at scale. While a few of the leading AI companies may have implemented other variations of such methods, DeepSeek’s blitz shows the potential of how front-end-edge AI can be developed on significantly fewer resources than previously thought.
With its open-source release, DeepSeek has laid the ground for cheaper AI development and potentially changed the course of the industry by making its methods known to the rest of the AI research world.
Its success supports the idea that only companies with billion-dollar budgets can compete to create the most sophisticated AI systems.
Cryptopolitan Academy: Coming Soon - A New Way to Earn Passive Income with DeFi in 2025. Learn More
