Re-math your weekend DeepSeek bill after the V4-Flash price cut

The cut is a timing discount

You open the dashboard after a long day and see a smaller number. DeepSeek's V4-Flash API price cut can reach 60 percent, with the new schedule starting September 10. The headline can read like a universal discount, but the bill does not work that way. Re-math the bill by token shape and time slot.

Before the August 17 price increase, the Flash API rate card was 0.02 yuan per million tokens for cache-hit input, 1 yuan for cache-miss input, and 2 yuan for output. The old card is useful for comparing a weekend bill before and after the rate-card switch. The updated rate card becomes active at 12:00 Beijing Time on September 10.

V4-Flash peak hours are Monday through Friday 9:00-12:00 and 14:00-18:00 Beijing Time, with all other times off-peak. For a weekend builder, the weekend is the default cheap window, while weekday runs need a closer look.

Peak hours are the real price switch

Off-peak, the card is 0.02 yuan per million tokens for cache-hit input, 1 yuan for cache-miss input, and 4 yuan for output. During peak hours, the card is 0.04 yuan per million tokens for cache-hit input, 2 yuan for cache-miss input, and 8 yuan for output. The spread is the main variable.

Long-prompt, short-answer jobs are dominated by the input line. Long-output jobs are dominated by the output line. A cache-hit run can cost very little, while a cache-miss run can add a larger input charge. A long report, a generated file, or a chatty agent can push that line higher than the input line. Cheaper tokens help only when the token shape matches the discount. Check the token breakdown before you assume the discount reached your job. A weekend job that is mostly writing may not match the headline discount.

Run a triage pass before you celebrate

• Classify each run. The finished state is a simple label on every job: cache-hit off-peak, cache-miss off-peak, or output/peak. The label tells you which line of the rate card is doing the work.
• Price the label in yuan first: use the matching cache-hit, cache-miss, or output line for the run's time slot. Then convert the yuan total to dollars at your card's rate, and keep both numbers in the same note. If a run has several parts, price each part separately and add the lines. The note should hold the label, the yuan total, and the dollar total. The dollar number is what your bank sees; the yuan number is what the API meter sees.
• Move the job if the time slot is the problem: if a weekday run falls in the peak window, shift it to the evening, weekend, or another off-peak window. A scheduler change is enough when it keeps the same prompt and moves the hour. A small shift can matter more than a big prompt change. Do not move a job just because it feels expensive. Move it when the hour, not the token shape, is the main cost driver.
• Cache the prompt if the input line is the problem: a stable system prompt, tool list, or context block can cut the input cost sharply. The cache pays off when the same long prefix is reused across runs instead of rebuilt.
• Kill the run if the output line is the problem and the output is not the product: if a job writes a long answer that nobody reads, the output line covers waste. A clean kill produces a shorter prompt, a stricter answer format, or a stopped job.

A side project can stay a side project. Keep the label, the yuan line, and the dollar line together, and the next decision is a quick check.

DeepSeek began internal testing of the model deepseek-v4.1-flash-expires-on-0910, with a 20-concurrent-request limit per account and temporary billing the same as V4 Flash. Do not point a weekend pipeline at it.