The Hangzhou-based company said in a statement on its developer platform that overall API prices would rise "in the near future", with detailed pricing to be announced later. It advised developers to plan their usage accordingly, warning that the increase would be substantial.
The announcement comes just one week after the launch of DeepSeek-V4-Flash-0731, a lightweight 284-billion-parameter version of the company's V4 model series.
Pressure on low-cost AI strategy
The planned price hike highlights the challenges DeepSeek faces in maintaining its reputation for offering high-performance AI models at exceptionally low prices amid intensifying global competition.
AI developer Michael Guo questioned the timing of the decision, arguing that newer models from rivals such as Meta and OpenAI now offer competitive performance and pricing.
Despite the planned increase, DeepSeek's latest model has gained widespread attention for delivering near frontier-level performance at a fraction of the cost of leading competitors.
Research firm Epoch AI ranked DeepSeek-V4-Flash-0731 as the world's second most capable open-weight AI model, behind only Beijing-based Moonshot AI's Kimi K3 model.
Much cheaper than rivals
According to benchmark firm Artificial Analysis, DeepSeek-V4-Flash-0731 currently costs around $0.03 per task, making it roughly 105 times cheaper than Anthropic's Claude Fable 5, which costs about $3.15 per task.
Epoch AI also noted that DeepSeek's API pricing remains significantly lower than comparable Chinese competitors, charging $0.14 per million input tokens and $0.28 per million output tokens, compared with Zhipu AI's GLM-5.2 pricing of $1.40 and $4.40, respectively.
The model has also become the fastest-growing AI model by token usage on Ollama, a platform that enables users to run AI models on local hardware. The company said it is expanding capacity across the United States and Europe to meet growing demand.
Efficiency behind low prices
Experts say DeepSeek's competitive pricing has been made possible by advances in model architecture and training techniques.
Zhenhui Jack Jiang, professor of Innovation and Information Management at the University of Hong Kong Business School, said the company's mixture-of-experts architecture activates only a small number of specialised components for each request, reducing computing costs while maintaining strong performance.
He added that DeepSeek also uses multi-teacher knowledge distillation, allowing a single model to efficiently learn capabilities from several specialised AI systems and reduce training costs.
While the company has not disclosed the exact timing or scale of the increase, the planned price revision marks a shift in DeepSeek's strategy as it balances rapid global demand with the costs of scaling its AI infrastructure.
