DeepSeek announces steep price increase due to rising demand for low-cost AI models
DeepSeek has announced that it will significantly increase the prices of its API services in the near future. According to the Chinese AI start-up, the new rates will be disclosed later, but users are advised to already take this into account in their usage.
The announcement comes a week after the introduction of DeepSeek-V4-Flash-0731, a lightweight AI model with 284 billion parameters. The model attracted worldwide attention because it delivers performance that comes close to the best AI models, while costing much less.
The price increase highlights the challenge DeepSeek faces. The company has built its reputation on very low prices, but must find a way to maintain that strategy in a market where competition is becoming increasingly fierce. AI developer Michael Guo wrote on X that a price increase at this moment could be risky, because new models from, among others, Meta and OpenAI in his view deliver comparable performance at competitive rates.
Research firm Epoch AI currently calls DeepSeek-V4-Flash-0731 the second most powerful open-weight AI model on the market, only surpassed by the Kimi K3 model from China’s Moonshot AI. Its performance is said to be somewhere between Anthropic’s Claude Opus 4.5 and Opus 4.6.
The model also strongly distinguishes itself in terms of price. According to Artificial Analysis, an average task with DeepSeek-V4-Flash-0731 costs about 3c (US), compared to $3.15 US dollars for Anthropic’s Claude Fable 5. That makes the DeepSeek model roughly 105 times cheaper.
Epoch AI points out that DeepSeek is not only cheaper than American competitors, but also much more affordable than comparable Chinese models. The current API prices are 14c per million input tokens and 0.28c per million output tokens, while competitor Zhipu AI charges $1.40 and $4.40 respectively.
The model’s popularity is growing rapidly. On the Ollama platform, which allows AI models to be run locally, DeepSeek-V4-Flash-0731 has, according to the company, become the fastest-growing model ever measured in terms of token usage. Ollama is therefore expanding capacity in the United States and Europe.
According to Professor Zhenhui Jack Jiang of the University of Hong Kong, who spoke to the South China Morning Post, DeepSeek owes its low prices to an efficient technical approach. The model uses a so-called mixture-of-experts architecture, in which only part of the network is activated for each task. As a result, less computing power and memory are required. In addition, DeepSeek uses a training method in which knowledge from multiple specialised AI models is combined into a single model, allowing it to be trained more efficiently.
Emerce
Subscribers 0
Fans 0
Followers 0
Followers