DeepSeek's New Peak-Hour API Pricing: How to Cut Your Bill in Half
DeepSeek quietly replaced flat API pricing with peak and off-peak rates on August 16. Here's the new price table, a timezone conversion guide, and how to time your calls to cut costs in half.
Contents5
What actually changed on August 16
DeepSeek quietly ended flat-rate API pricing at 16:00 UTC on August 16, 2026. In its place: peak and off-peak rates that swing by as much as 2x depending on when you send the request. This is a separate change from the new V4-Pro model DeepSeek shipped the same week, and it applies across the V4 family, including V4-Flash-0731, the model most agent builders default to for cost reasons.
If you're calling the DeepSeek API on a schedule, or building anything that can tolerate a delay, this is worth 15 minutes of your time. The difference between calling at the right hour and the wrong one is real money at scale.
The new price table
Peak hours are defined in Beijing time: 9 a.m. to 12 p.m. and 2 p.m. to 6 p.m. Off-peak rates apply the rest of the day and are roughly half of peak.
| Model | Metric | Old flat rate | New peak rate | New off-peak rate |
|---|---|---|---|---|
| V4-Flash-0731 | Output | $0.28 / M tokens | $1.32 / M tokens | $0.66 / M tokens |
| V4-Pro | Output | $0.84 / M tokens (converted) | $3.96 / M tokens | about $1.98 / M tokens |
| V4-Pro | Cached input | 0.025 yuan / M | 0.3 yuan (about $0.04) / M | about 0.15 yuan / M |
| V4-Pro | Uncached input | 3 yuan / M | 9 yuan / M | about 4.5 yuan / M |
Two things jump out. First, even the "cheap" off-peak rate on V4-Flash output is more than double the old flat price. This isn't a discount window, it's a smaller markup window. Second, the peak multiplier on V4-Pro cached input is enormous in relative terms, 12x, even though the absolute dollar amount is tiny. Don't let the percentage headline distract you from what it actually costs on your bill.
If you already read our piece on [DeepSeek's V4-Pro launch and the 1,100% price hike](https://questloops.com/blog/deepseek-just-raised-api-prices-by-up-to-1-100-what-changed-and-what-to-do-about-it), this peak/off-peak split is the mechanism behind those headline numbers. The 1,100% figure is the peak-hour, worst-case comparison.
Converting Beijing peak hours to your clock
Beijing time is UTC+8 with no daylight saving, which makes the conversion static year-round. Here's when peak pricing actually hits if you're calling the API from outside China:
| Region | Peak window 1 | Peak window 2 |
|---|---|---|
| UTC | 01:00 to 04:00 | 06:00 to 10:00 |
| US Eastern (EDT) | 9:00 p.m. to midnight (prev. day) | 2:00 a.m. to 6:00 a.m. |
| US Pacific (PDT) | 6:00 p.m. to 9:00 p.m. (prev. day) | 11:00 p.m. to 3:00 a.m. |
| UK/Western Europe (BST) | 2:00 a.m. to 5:00 a.m. | 7:00 a.m. to 11:00 a.m. |
| India (IST) | 6:30 a.m. to 9:30 a.m. | 11:30 a.m. to 3:30 p.m. |
The useful takeaway for US-based teams: DeepSeek's peak windows land mostly during US nighttime hours. If your workload already runs on a normal 9-to-5 US schedule, you're closer to off-peak by accident than you might think. The exception is anything that fires overnight batch jobs, retraining pipelines, scheduled scraping, nightly reports. Those are the calls most likely to walk straight into peak pricing without anyone noticing.
For India and Western Europe, it's the opposite problem. Normal business hours overlap with Beijing's peak windows almost exactly, so there's no free lunch from timezone alone.
How to actually act on this
Tag your DeepSeek calls by urgency first. Anything that isn't user-facing, like batch summarization, embeddings refreshes, or eval runs, should get queued and released during the off-peak window for your region rather than firing whenever a cron job happens to wake up.
Don't assume a routing gateway handles this for you. Tools like OpenRouter pass through DeepSeek's live pricing as-is; they don't arbitrage the peak/off-peak gap unless you build the scheduling logic yourself.
Recheck your cost dashboards weekly for now. DeepSeek has changed V4 pricing twice in the past month, so a number you verified two weeks ago may already be wrong.
If token costs are a growing line item beyond just DeepSeek, our guide on [cutting your AI agent's API bill](https://questloops.com/blog/how-to-cut-your-ai-agent-s-api-bill-token-compression-smart-routing-and-when-a-gateway-pays-for-itself) covers compression and routing tactics that stack on top of timing your calls.
The bottom line
Off-peak DeepSeek pricing is not a discount, it's damage control. Shift your workload into the off-peak window and you'll roughly halve the increase. But go in knowing the baseline moved for everyone, and the timezone math genuinely favors US business hours more than most other regions.