DeepSeek's published price table now treats both weekend days as entirely off-peak. Peak pricing applies only from 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday; everything else, Saturday and Sunday included, bills at half the peak rate. The change took effect at 00:00 Beijing time on 23 August — 16:00 UTC on 22 August.
The current table
For deepseek-v4-pro, output runs $1.98 per million tokens off-peak against $3.96 at peak; cache-miss input is $0.66 against $1.32, and cache-hit input $0.022 against $0.044. For deepseek-v4-flash and the experimental vision variant, output is $0.66 against $1.32 and cache-miss input $0.22 against $0.44. The rule is uniform: off-peak is exactly half.
The distortion: a concession inside an increase, not a discount
Read on its own, weekend off-peak pricing looks like DeepSeek cutting prices. Set against the previous table it is the opposite. The peak/off-peak structure itself only took effect at 16:00 UTC on 16 August, and it raised prices sharply. V4-Pro output was a flat $0.87 per million tokens before that date, so the new weekend “discount” rate of $1.98 is about 2.3 times the old price. V4-Flash output went from a flat $0.28 to $0.66 off-peak, roughly 2.4 times. Nobody is paying less this weekend than they paid two weeks ago; they are paying less than they would have under the structure introduced six days earlier.
What actually changed on 22 August
Only the calendar. The rates did not move — the definition of when peak applies did. Under the 16 August rules, weekend hours falling inside 01:00–04:00 or 06:00–10:00 UTC billed at peak. They no longer do. For a workload running continuously, the saving is confined to those seven hours per weekend day.
Cache pricing is where the real spread sits
The widest ratio in the table is not between peak and off-peak but between cache hit and cache miss: $0.022 against $0.66 per million input tokens on V4-Pro, a factor of thirty. Prompt structure that preserves cache hits moves the bill considerably further than choosing which day to run.
Who the change is aimed at
Batch and evaluation workloads that can be scheduled. A team able to defer a large offline job to Saturday now pays half rate across the entire run rather than across part of it, which makes weekend scheduling a real lever — just a smaller one than “prices fell” would imply.