DeepSeek's New Rule: Weekends Now Off-Peak Rate
DeepSeek is moving fast on this one. It only rolled out its peak/off-peak dual-tier billing system on August 17, and less than a week later, an official email landed: starting 00:00 Beijing time on Sunday, August 23, weekends will no longer distinguish between peak and off-peak hours — every call all day will be billed at the off-peak rate. In other words, the peak-hour pricing logic now only applies on the five weekdays; on Saturday and Sunday, whatever time you call the API, you get the cheaper rate.
What Actually Changed
According to DeepSeek’s official notice, the scope of this change is simple: weekdays (Monday through Friday, Beijing time) keep the existing peak/off-peak tiered billing rules unchanged; weekends (Saturday and Sunday, Beijing time) drop the peak/off-peak distinction entirely for the whole day, with every call billed at the off-peak rate. Compared to the rules that just went live on August 17 — peak hours set at 9:00–12:00 and 14:00–18:00 Beijing time, with off-peak priced at half the peak rate — this change effectively “locks” both weekend days into the cheaper off-peak tier, so calling the API at 2 p.m. on a Saturday no longer gets counted as peak-hour usage.
DeepSeek’s stated reasoning is to give developers “more flexibility in scheduling weekend usage without worrying about peak-hour costs,” while also noting that the change helps the company “balance overall computing loads” — in plain terms, it wants to nudge workloads that would otherwise get stuck behind weekday peak pricing, but aren’t urgent, toward the weekend when servers are relatively idle.
Effective Date and Transition Terms
The email is explicit about the effective moment: 00:00 Beijing time on Sunday, August 23, 2026, at which point the system will automatically apply the new billing standard; any charges incurred before that moment will still be settled under the original billing rules, with no retroactive application of the new ones. DeepSeek also included its standard terms notice at the end of the email: continuing to use the service after the billing adjustment is considered acceptance of the new terms, and if you disagree, you may cancel your service and apply for a refund.
Worth noting: Taiwan and Beijing share the same UTC+8 time zone, so this “00:00 Beijing time” lands at exactly the same moment for developers in Taiwan — no time-zone conversion needed.
The Full Pricing Table: Three Models
DeepSeek’s official pricing page currently lists three models, with the following pricing structure (in RMB per million tokens):
| Model | Time slot | Input (cache hit) | Input (cache miss) | Output | Concurrency limit |
|---|---|---|---|---|---|
| deepseek-v4-flash | Off-peak | ¥0.05 | ¥1.5 | ¥4.5 | 2500 |
| deepseek-v4-flash | Peak | ¥0.10 | ¥3.0 | ¥9.0 | 2500 |
| deepseek-v4-pro | Off-peak | ¥0.15 | ¥4.5 | ¥13.5 | 500 |
| deepseek-v4-pro | Peak | ¥0.30 | ¥9.0 | ¥27.0 | 500 |
| deepseek-v4-flash-vision-exp | Off-peak | ¥0.05 | ¥1.5 | ¥4.5 | 2500 |
| deepseek-v4-flash-vision-exp | Peak | ¥0.10 | ¥3.0 | ¥9.0 | 2500 |
The three models map to underlying versions DeepSeek-V4-Flash-0731, DeepSeek-V4-Pro-0813, and the newly-appeared experimental vision model DeepSeek-V4-Flash-Vision-Exp — this model wasn’t yet on the official pricing page when we covered the August 17 price change. Its pricing structure and concurrency limit are identical to the regular flash model, with the added capability of processing image inputs, which get converted into a corresponding token count based on image dimensions and billed alongside text tokens. All three models share a 1M context length and a maximum output length of 384K, and all support JSON output, tool calls, the Responses API, and Anthropic API compatibility; the main difference is FIM (fill-in-the-middle) completion — flash and pro support it only in non-thinking mode, while vision-exp doesn’t support it at all.
What This Actually Means for Developers
If your workflow already has some flexibility and isn’t racing against weekday peak hours, this change amounts to a free savings window: shift any deferrable batch work — data preprocessing, offline content generation, model benchmarking — to Saturday or Sunday, and no matter what time you call the API, you get the off-peak rate, without needing to deliberately dodge the 9-to-12 and 2-to-6 peak windows the way you would on weekdays. For teams already in the habit of concentrating heavy compute on weekends, this is close to a pure win. And if your use case genuinely needs high-frequency calls during weekend peak hours — say, a consumer-facing product where weekend traffic actually spikes — this change still works in your favor, since there’s no more “peak surcharge” to worry about on weekends at all. Overall, this is DeepSeek’s first user-friendly fine-tuning of its billing rules since the August 17 price increase.
