DeepSeek Off-Peak Pricing Now Applies All Weekend: What Heavy Users Save

DeepSeek just handed heavy AI users a new way to cut the bill, and it only works on Saturdays and Sundays. Effective 00:00 Beijing Time on Sunday, August 23, DeepSeek adjusted its peak and off-peak billing rules so that off-peak rates now apply throughout the day on weekends. Every hour of Saturday and Sunday is billed at the off-peak price, not just the narrow weekday windows that were off-peak before.

For anyone running batch jobs, retry loops, or heavy agentic pipelines through the DeepSeek API, this is the most directly actionable pricing change of the week. Off-peak is exactly half the peak rate. Move your most token-hungry work to the weekend and the input and output line items drop by 50 percent. Here is the math, the timing, and the routing playbook to capture it.

What the DeepSeek weekend off-peak change means

Before this change, off-peak pricing applied only outside the weekday peak windows of 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday. The weekend was an ambiguous gray zone. Now DeepSeek has made it explicit: Saturdays and Sundays, by Beijing Time, are off-peak all day.

The effect is a clean doubling of your off-peak capacity. On a normal weekday you get roughly 19 of 24 hours at off-peak rates. On a weekend you get all 24. For a provider whose off-peak rate is a flat 50 percent discount, that is effectively a second, fully discounted weekend shift you can point your heaviest workloads at.

The current per-million-token rates that apply all weekend:

• deepseek-v4-flash: cache-hit input $0.007, cache-miss input $0.22, output $0.66
• deepseek-v4-pro: cache-hit input $0.022, cache-miss input $0.66, output $1.98

At peak those same numbers roughly double. V4-Flash output jumps from $0.66 to $1.32, and V4-Pro output from $1.98 to $3.96. The weekend discount is not a trick or a two-hour lucky window. It is the entire two-day calendar, and it is worth planning around.

Why a full weekend of off-peak rates is a big deal

The weekday off-peak windows are awkward for most teams. The 04:00 to 06:00 UTC gap is the middle of the night for North America and the very early hours for Europe. The 10:00 to 01:00 UTC stretch overlaps the European workday and the start of the US East Coast morning, which is exactly when most people are actively using their agents and want fast, uncongested responses rather than deep discounts.

Weekends remove that conflict. When nobody is actively driving the agents, you do not need low latency. You need low cost. An all-day off-peak weekend lets you schedule the expensive, output-heavy work for the hours that have no other users competing with you anyway. That is the whole point: you are paying the off-peak price for traffic no one else wants.

For a heavy user moving even 10 million output tokens a month through V4-Pro, the difference between all-weekend off-peak and weekday peak routing is roughly $19,800 a month on the output line alone before caching. Even a fraction of that shift is not noise. It is real money on the token bill.

How to route your weekend batch work

The old DeepSeek playbook told you to chase the 04:00 to 06:00 and 10:00 to 01:00 UTC weekday gaps. That still works, but the better play now is to concentrate the truly heavy work on Saturday and Sunday. Here is the routing order that captures the most:

• Schedule non-interactive batch jobs, data pipelines, and retry loops for Saturday and Sunday. Those are your biggest off-peak winners because they can tolerate multi-hour queues.
• Keep the weekday off-peak gaps for the smaller, time-sensitive batches you cannot hold to the weekend. They still get the 50 percent discount.
• Reserve peak hours for genuinely interactive work, when a human is waiting on a response. That keeps cost down on everything that is not latency-critical.
• Watch timezone math if your team spans Beijing. The weekend boundary is set in Beijing Time, so a Friday evening in North America is already Saturday in Beijing and already off-peak.

Caching still matters more on weekends

The weekend discount changes the price per token, but it does not change the relative value of cache hits, and that is where you can double up the savings. Cache-hit input is between 30 and 60 times cheaper than cache-miss input on the same model. At V4-Flash that is $0.007 versus $0.22 per million. At V4-Pro it is $0.022 versus $0.66.

Over a big weekend batch, a low cache-hit rate can quietly negate the off-peak discount. Keep shared prefixes, system prompts, and repeated context cached so your input side stays near the cache-hit floor. The weekend pricing is the headline, but cache discipline is what keeps it a clean win instead of a wash.

What has not changed

DeepSeek has not changed the flat 50 percent off-peak discount, the weekday peak windows, or the per-model rates themselves. Those are the same numbers that shipped on August 16 with the V4-Pro general availability release. What changed is only the calendar on which off-peak fully applies. That means you should not re-baseline your cost model, because the rates are identical. You should re-baseline your schedule, because there is now a two-day window every week where the whole day is discounted.

The broad price picture from last month still holds. DeepSeek off-peak is no longer the ultra-cheap floor it was before the August increase, but the weekend discount is a genuine new lever that restores part of that advantage for anyone willing to shift workload around a calendar.

The bottom line for heavy users

DeepSeek off-peak pricing now covers every hour of Saturday and Sunday, and that is a real, budget-sized win if you act on it. Shift the token-heavy batch, agentic, and retry workloads to the weekend where no one else is competing for capacity anyway, keep those weekday off-peak gaps for the smaller time-sensitive runs, and protect the input side with aggressive caching. The per-token rates did not move. The opportunity did, and it opens for two days every week starting now.