DeepSeek's 'Off-Peak Discount' Kicks In Tomorrow. Every Published Rate Is Still Above What You Pay Today.
At 16:00 UTC on August 16, DeepSeek moves its API from flat per-token pricing to a peak/off-peak schedule. Today, V4-Pro costs $0.435 per million input tokens and $0.87 per million output tokens, cache misses included. Starting tomorrow, the cheapest hour you can buy is $0.66 input and $1.98 output. That's not a discount tier sitting below today's price. It's a floor sitting above it, and DeepSeek is calling the whole thing an off-peak rate structure instead of what it actually is: a price increase with a scheduling gimmick attached.
I run a chunk of my own tooling through DeepSeek specifically because it's been the cheapest credible option for batch work: nightly summarization, log triage, the stuff that doesn't need to happen in real time. If you've built anything similar, this update is worth ten minutes of your attention before it lands, not after.
The actual numbers, side by side
DeepSeek's own pricing announcement lays out two new bands. Peak hours are 01:00-04:00 UTC and 06:00-10:00 UTC. Everything else counts as off-peak. For V4-Pro, peak pricing is $1.32 input and $3.96 output per million tokens, off-peak is $0.66 and $1.98. For V4-Flash, peak is $0.44 and $1.32, off-peak is $0.22 and $0.66.
Line those up against today's flat rate and the gap is stark. V4-Pro input goes from $0.435 to a $0.66 floor, a 52% jump on the cheapest hour available. Output goes from $0.87 to $1.98, up 127%. Peak-hour output at $3.96 is more than four and a half times what you pay right now. Off-peak isn't a discount relative to today. It's the new baseline, and peak is the penalty for using the API when everyone else does too.
Why this reads as a discount when it isn't
The framing does real work here. "Off-peak pricing" implies two tiers where one is cheaper than some default and the other is more expensive, and most people's mental model fills in "cheaper than what I pay now" without checking. DeepSeek isn't lying about the numbers, both bands are published clearly, but putting "off-peak" and "discount" in the same sentence when the off-peak rate is still 52-127% above the current flat rate is a framing choice, not an accident. Companies restructure pricing to increase revenue per token fairly often. What's unusual is dressing an increase up in language that reads as a savings opportunity to anyone who doesn't do the arithmetic.
This lands three days after DeepSeek's quiet V4-Pro GA release, which I covered here without much fanfare because DeepSeek didn't put out a blog post about it either. The pattern now looks less like a coincidence and more like a company that ships pricing changes with minimal noise on both ends, the capability bump and the bill increase.
What this actually costs you
Take a workload that runs 500 million input tokens and 100 million output tokens a month, off-peak only, which is a realistic shape for scheduled batch jobs. Today that's $217.50 input plus $87 output, $304.50 total. Under the new off-peak rate it's $330 input plus $198 output, $528 total. That's a 73% increase for doing nothing differently, just because the calendar flipped to August 16.
If any of that workload spills into peak hours (01:00-04:00 and 06:00-10:00 UTC cover a meaningful chunk of the day, not a narrow overnight window), the number gets worse fast. Peak-hour output at $3.96 per million tokens means a job that used to cost you under a dollar could cost four times that if it happens to land in the wrong four-hour block.
The self-host math shifts too
DeepSeek's earlier "permanent 75% price cut" was the thing that made API pricing beat self-hosting for a lot of solo operators running V4-Pro-class workloads. That calculus was built on $0.435/$0.87. At $0.66-$1.32 input and $1.98-$3.96 output, the crossover point where renting GPU hours or running a smaller open-weight model locally beats the API moves closer. I'm not saying self-hosting suddenly wins for everyone, most solo operators don't want to manage GPU infrastructure for a workload that used to cost $300 a month. But the margin that made "just use the API" an easy default just got thinner, and for higher-volume batch pipelines it might have flipped outright.
What I'd actually do
Pull your last 30 days of DeepSeek usage before August 16 and re-run it against both new bands. If your batch jobs cluster inside the 01:00-04:00 or 06:00-10:00 UTC windows, even by accident because that's when your cron scheduler happens to fire, move them. A five-line change to a cron expression that shifts a job from 07:00 UTC to 11:00 UTC is the cheapest optimization available to you this week, and it's the difference between the off-peak rate and the peak rate.
Beyond that, actually compare the new off-peak floor against V4-Flash for anything that doesn't need V4-Pro's full capability, and against a self-hosted option like Qwen3.8-27B (open-weighted this week, runs on a single high-end GPU) if your volume is high enough that the delta compounds into real money. Where I could be wrong: DeepSeek has cut prices aggressively and repeatedly through 2026, and it's plausible this "increase" is a temporary rebalancing ahead of another cut once the new billing infrastructure is in place. I wouldn't bet a migration plan on that, but if you're mid-project and the increase is survivable, watching for a few weeks before ripping out your integration isn't unreasonable either.
Author
Lukas
@lukcombinator