Chinese artificial intelligence (AI) startup DeepSeek officially launched its latest model, "V4 Pro," and raised API prices by up to 12 times. As DeepSeek, which had touted industry-low prices, sharply increased prices, some developers said "DeepSeek's unique cost-effectiveness is gone," while local media analyzed that the new model's performance will ultimately determine the real value for money.
According to China Business News, DeepSeek announced the official launch of V4 Pro on the evening of the 13th and released a revamped API pricing plan. The new prices will take effect from 12 a.m. on the 17th (local time).
The biggest feature of the new pricing overhaul is the introduction of peak and off-peak plans. 9 a.m.–12 p.m. and 2–6 p.m. are designated as peak hours, and the remaining hours are classified as off-peak, with different service prices. During peak hours, API prices are 9 yuan per 1 million tokens for input and 27 yuan for output, up threefold and 4.5 times, respectively, from before.
In particular, based on peak hours, the price for "cache hit" input that reuses previously entered data rose from 0.025 yuan to 0.3 yuan per 1 million tokens, a 12-fold increase. During off-peak hours, half the price applies. DeepSeek said it aims to allocate resources more rationally and advised users to adjust their working hours based on actual usage.
The price increase does not apply to general users. Website and mobile application (app) users can use the official V4 Pro version for free through "expert mode."
Those facing greater expense burdens are developers and corporate customers who connect DeepSeek models to their own services through the API. Among developers, some say the scale of the price hike is excessive.
One developer told local media, "DeepSeek is no longer cheap. Its prices are now similar to peer models," adding, "In particular, the large increase during peak hours has erased DeepSeek's unique cost-effectiveness." Some also expect that as Ascend, a Chinese AI Semiconductor chip by Huawei, spreads quickly in the market, DeepSeek will lower prices again.
However, the new model's performance does not appear to be overwhelmingly superior to other AI models. At Artificial Analysis, an AI model evaluation agency, V4 Pro's intelligence score was 53, ranking ninth overall, while on Arena, V4 Pro Max ranked around eighth overall and second among open-source models after Moonshot's "Kimi K3 Max." However, because the results are not yet based on large-scale user voting, the rankings could change.
China Business News said, "With this price increase, DeepSeek is no longer inexpensive, and the actual value for money ultimately depends on performance," adding, "Whether the new model can enable developers to handle more tasks with fewer tokens in real production environments will likely determine whether developers can accept the higher expense."