DeepSeek’s time-dependent tariff needs both a calendar and a token breakdown. The largest increase on one price line is not automatically the increase in a customer’s total bill.
The current tariff and the original announcement
DeepSeek's August 13 release notice sets the new pricing's start at 16:00 UTC on August 16. The current official table lists V4 Pro, per million tokens: cached input $0.022 off-peak / $0.044 peak, uncached input $0.66 / $1.32, and output $1.98 / $3.96.
At this review, peak windows are 01:00-04:00 and 06:00-10:00 UTC, Monday through Friday. All other hours are off-peak. The earlier article described daily peaks without the current weekday restriction. Use the live table rather than a permanently hard-coded calendar copied from launch coverage.
Price a mixed request
Our fictional request contains 80,000 cached input tokens, 20,000 uncached input tokens and 5,000 output tokens. Off-peak, the components cost $0.00176, $0.0132 and $0.0099: $0.02486 total. At peak, the same mixture costs $0.04972. These are token charges only, not a measured account invoice.
The cached portion is 80% of input volume but only about 7.08% of this off-peak bill. Output accounts for roughly 39.82%, and uncached input 53.10%. This is why a spectacular percentage change in cache pricing need not dominate a real workload. Conversely, a different mixture can make caching much more significant.
The diagram shows both total and component widths from the same calculation. It assumes the stated cache hits actually occur; resending text does not, by itself, prove it was billed as cached.
Instrument before rescheduling
Record model, returned usage categories, request timing and billed amount. Scheduling a job in an off-peak window does not guarantee all its retries or later calls stay there. Keep UTC as the calculation basis and convert for display with the correct date-aware time zone.
Move delay-tolerant work only when its deadline and service behavior allow it. Compare the savings with extra queueing, retries and operational complexity. A time-based tariff encourages load shifting, but it does not establish the provider's private capacity, margins or fundraising motives. Those explanations in the original article were inference, not documented facts.
Update weekday-only peak schedule from current official docs; replace maximum line-item percentage with a worked token-mix bill; do not infer private capacity economics.