At 00:00 on August 17, DeepSeek officially rolled out its new time‑of‑day pricing mechanism for API services, making it one of China’s first large‑model vendors adopting time‑segment billing. This marks a shift for the domestic large‑model industry from simple price competition toward refined commercial operation. Under the new rules, a full day is divided into peak, flat‑peak and off‑peak periods with differentiated unit prices for API calls. Costs drop substantially during late‑night off‑peak hours, helping corporate clients cut computing‑power procurement expenses.

Previously, most domestic LLM APIs adopted flat‑rate pricing, charging the same price per thousand tokens regardless of day‑time or night‑time usage. Inspired by power‑grid scheduling logic, time‑of‑day pricing addresses real‑world resource constraints. GPU clusters face tight resource pressure in daytime, while server utilization rates drop significantly at night. Price incentives guide clients to shift non‑urgent inference workloads to nighttime hours, raising overall GPU utilization and easing hardware‑depreciation pressure for AI developers.

For enterprise users, non‑real‑time workloads such as batch document parsing, data cleaning and offline fine‑tuning can be scheduled during off‑peak windows to greatly reduce AI‑related expenditure. By contrast, low‑latency services including chatbots and real‑time retrieval‑augmented generation still need to run during peak hours. Industry analysts point out that given continuously high hardware investment for large models, endless price cuts are not commercially viable. Time‑of‑day pricing represents a meaningful attempt to balance computing‑power supply and demand and improve gross margins for AI companies. It is widely expected that more domestic LLM players will launch similar pricing schemes in the near future.