DeepSeek Raises API Prices — Batch Jobs Halve It
The rise is real, but half of it is optional. Peak is only seven hours a day.
Founder, Automation Squad ·
The short answer
DeepSeek split its pricing in two at 16:00 UTC on August 16, 2026. Peak is 01:00–04:00 and 06:00–10:00 UTC. Everything else is off-peak, at half. V4-Pro output was $0.87 per million tokens; it is now $3.96 peak, $1.98 off-peak. Shift a job out of those seven hours and you pay the lower number.
The new DeepSeek prices, and the jobs worth moving now
These rates are live now. Working through the scheduling pass below takes about a quarter of an hour and shows up on your next bill.
List every scheduled job that calls DeepSeek, with its UTC start hour
Cron entries, n8n schedule triggers, Make scenarios, GitHub Actions. Write the hour in UTC, not local time — this is where the mistakes happen.
Flag anything starting 01:00–04:00 or 06:00–10:00 UTC
Those are the peak windows. A job in there pays double for no benefit unless a human is genuinely waiting on the output.
Move the flagged jobs to 11:00 UTC or later
11:00–01:00 UTC is a continuous fourteen-hour off-peak block, so almost anything overnight can move without a fight. Change the cron expression, nothing else.
Check whether your prompts have a stable prefix
Cache-hit input on V4-Pro is roughly thirty times cheaper than cache-miss. Put the unchanging system instructions first and the variable content last, so the prefix can be cached across calls.
Re-check your spend after a week
Compare the same weekday before and after. If the bill did not move, something is still firing in a peak window — usually a retry or a job you forgot you had.
| Per million tokens | Was | Off-peak | Peak |
|---|---|---|---|
| V4-Pro — output | $0.87 | $1.98 | $3.96 |
| V4-Pro — input, cache miss | $0.435 | $0.66 | $1.32 |
| V4-Pro — input, cache hit | $0.003625 | $0.022 | $0.044 |
| V4-Flash — output | $0.28 | $0.66 | $1.32 |
| V4-Flash — input, cache miss | $0.14 | $0.22 | $0.44 |
| V4-Flash — input, cache hit | $0.0028 | $0.007 | $0.014 |
Take it with you
DEEPSEEK REPRICING — effective 16:00 UTC, Aug 16 2026
Source: https://automationsquad.com/news/deepseek-api-price-increase-august-2026/
PEAK HOURS (UTC): 01:00-04:00 and 06:00-10:00
OFF-PEAK: every other hour — half the peak rate
OFF-PEAK BLOCK: 11:00 UTC -> 01:00 UTC is 14 continuous hours
PRICES PER MILLION TOKENS was off-peak peak
V4-Pro output $0.87 $1.98 $3.96
V4-Pro input cache miss $0.435 $0.66 $1.32
V4-Pro input cache hit $0.003625 $0.022 $0.044
V4-Flash output $0.28 $0.66 $1.32
V4-Flash input cache miss $0.14 $0.22 $0.44
V4-Flash input cache hit $0.0028 $0.007 $0.014
ACTIONS
[ ] List every scheduled DeepSeek job + its UTC start hour
[ ] Flag anything in 01:00-04:00 or 06:00-10:00 UTC
[ ] Move flagged jobs to 11:00 UTC or later
[ ] Put unchanging system instructions FIRST so the prefix caches
(cache hit is ~30x cheaper than cache miss on V4-Pro)
[ ] Re-check spend after one week, same weekdayDeepSeek announced the change on August 13, 2026, and it took effect at 16:00 UTC on August 16. The new structure splits every rate into peak and off-peak, with peak roughly double off-peak across the board.
The facts, from DeepSeek's own pricing documentation: DeepSeek-V4-Pro output moves from $0.87 per million tokens to $3.96 at peak and $1.98 off-peak. Its cache-miss input goes from $0.435 to $1.32 peak and $0.66 off-peak. DeepSeek-V4-Flash output moves from $0.28 to $1.32 peak and $0.66 off-peak. Cache-hit input, the cheapest tier, goes from $0.003625 to $0.044 peak and $0.022 off-peak on V4-Pro. Peak hours are 01:00–04:00 and 06:00–10:00 UTC; every other hour of the day bills at the off-peak rate.
Automation Squad's take: the headline number — a 4.5x jump on V4-Pro output — is the one that will travel, and it is accurate for anyone who runs at peak. But peak is seven hours out of twenty-four, and the workloads most exposed to this increase are exactly the ones that do not care what time they run: overnight summarization, bulk classification, document extraction, embedding backfills. If your job has no human waiting on it, the price rise is half what the headline says, and the change costs you a cron expression. The genuinely painful case is interactive traffic during European or Asian business hours, which lands squarely in the peak windows and cannot be moved.
Run this now: list every scheduled job you have that calls DeepSeek and write the UTC hour it fires next to each one. Anything currently starting between 01:00–04:00 or 06:00–10:00 UTC is paying double for no benefit — move it to 11:00 UTC or later and re-check the cost after a week. While you are in there, note whether the job re-sends the same system prompt every call: cache-hit input is still roughly thirty times cheaper than cache-miss input on V4-Pro, so a stable prompt prefix is worth more now than it was last week.
By the numbers
The data behind the story — figures from cited sources, with our own analysis labelled.
| V4-Pro output — was | $0.87 |
| V4-Pro output — off-peak | $1.98 |
| V4-Pro output — peak | $3.96 |
| V4-Pro input, cache miss — was | $0.435 |
| V4-Pro input, cache miss — off-peak / peak | $0.66 / $1.32 |
| V4-Pro input, cache hit — off-peak / peak | $0.022 / $0.044 |
| V4-Flash output — was | $0.28 |
| V4-Flash output — off-peak / peak | $0.66 / $1.32 |
| V4-Flash input, cache miss — off-peak / peak | $0.22 / $0.44 |
Effective 16:00 UTC, August 16, 2026. Peak hours are 01:00–04:00 and 06:00–10:00 UTC; all other hours bill off-peak.
Questions people are asking
- What are DeepSeek's peak hours?
- Two blocks: 01:00–04:00 and 06:00–10:00 UTC. Seven hours in total. Everything outside them bills at half, and because the blocks sit close together you get one unbroken off-peak stretch from 11:00 UTC round to 01:00 UTC — fourteen hours to put a batch job in.
- How much did DeepSeek prices actually go up?
- Depends when you run. On V4-Pro output the peak rate went $0.87 to $3.96, so about 4.5x. Off-peak it went to $1.98, about 2.3x. Coverage quoting one enormous number is quoting the peak rate and leaving out that peak is seven hours.
- Does this affect interactive use or just batch jobs?
- Both are billed the same way, but only one is fixable. If nobody is sitting there waiting on the output, the job can move, and moving it costs you one edit to a cron expression. Live traffic is the painful case: European and Asian business hours sit right on top of the peak windows, and you cannot ask customers to come back at 11:00 UTC.
- What is cache-hit input and why does it matter more now?
- If the opening of your prompt matches a previous call word for word, DeepSeek serves that part from cache — on V4-Pro, at roughly a thirtieth of the cache-miss price. So the ordering inside your prompt is worth money: put the instructions that never change first, and the variable content last. That was always true. At the new rates it is worth more.
Sources
Further reading
- DeepSeek API — Models & Pricing — The live rate card, including the cache-hit tier most people forget to design for.
Last checked August 16, 2026 against the primary sources above, by Automation Squad Research. Spot an error? [email protected].