DeepSeek cuts API prices in off-peak times, between 00:30 and 08:30 in Beijing, by 75% for R1 and 50% for V3; the times are during US and Europe daylight hours
Context & Ripple Effects
The discount follows a brief capacity constraint: DeepSeek had suspended API-credit top-ups after demand overwhelmed its servers, then reopened top-ups as Chinese AI competition intensified the day before this move. Time-based pricing gives it a way to direct demand into lower-load windows rather than simply ration access.
The pricing approach also foreshadows DeepSeek's later decision to make a 75% V4 Pro API price reduction permanent, suggesting that aggressive API pricing became a recurring part of its model distribution strategy.
First-order effects
- Developers able to schedule inference in the stated Beijing off-peak window receive materially lower R1 and V3 API costs, with the largest immediate benefit for R1 workloads.
- DeepSeek can use the discount to shift request volume toward less-contended capacity while competing for developers active during US and European daytime.
Second-order effects
- Rival model providers serving the same developer hours face pressure to match either effective token prices or availability, especially for workloads that can be queued or scheduled.
- Customers may redesign batch jobs, evaluation runs, and other delay-tolerant inference around discount windows, making time-of-day a purchasing variable alongside model quality and headline price.
Third-order effects
- If sustained, capacity-aware pricing could make AI API markets more operationally differentiated: buyers will optimize for when inference runs, not only which model they use.
- The episode points toward inference economics becoming a competitive lever, where providers use utilization and pricing together to convert available compute into developer adoption.
The trend: AI model vendors are increasingly turning compute utilization into a pricing tool as they compete to lower the effective cost of inference.