Off-Peak AI API Pricing Explained: 2026 Guide
The Short Answer
Off-peak pricing means the same model costs different amounts at different times of day. It arrived in the mainstream on August 16, 2026, when DeepSeek moved its V4 family from flat rates to a peak/off-peak schedule where off-peak is exactly half of peak.
If any part of your workload can wait a few hours, this is the cheapest optimisation available — cheaper than switching models, and it costs you nothing in quality.
How It Works
Providers price compute the way electricity utilities do: demand is not evenly distributed across the day, capacity is fixed in the short run, so price is used to flatten the curve.
DeepSeek V4 (from August 16, 2026):
| Off-peak | Peak | |
|---|---|---|
| V4-Flash input / output | $0.22 / $0.66 | $0.44 / $1.32 |
| V4-Pro input / output | $0.66 / $1.98 | $1.32 / $3.96 |
Peak hours: 01:00-04:00 and 06:00-10:00 UTC. That is seven hours. The other 17 hours are half price.
Z.ai GLM Coding Plan: off-peak usage — outside 14:00-18:00 UTC+8 on weekdays — consumes 50% fewer points. Same idea, different unit.
The Timezone Lottery
This is the part that decides whether off-peak pricing is a gift or a tax, and most teams never check it.
DeepSeek’s peak window in local time:
| Location | Peak hours locally | Business day impact |
|---|---|---|
| US Pacific (UTC-7) | 18:00-21:00, 23:00-03:00 | None — fully off-peak |
| US Eastern (UTC-4) | 21:00-00:00, 02:00-06:00 | None — fully off-peak |
| London (UTC+1) | 02:00-05:00, 07:00-11:00 | Morning hit |
| Berlin (UTC+2) | 03:00-06:00, 08:00-12:00 | Morning hit |
| Tallinn (UTC+3) | 04:00-07:00, 09:00-13:00 | Morning hit |
| India (UTC+5:30) | 06:30-09:30, 11:30-15:30 | Most of the day |
| China (UTC+8) | 09:00-12:00, 14:00-18:00 | Entire business day |
An American team using DeepSeek pays off-peak rates without changing anything. A Beijing team pays double for the same work. When a provider’s peak window matches its home market’s working hours, the schedule is effectively an export discount.
Action: convert your provider’s peak window into your local time and mark it on a calendar. Until you do, you are guessing at half your unit economics.
What To Move Off-Peak
Sort your workloads by whether a human is waiting:
Move it — nobody is waiting:
- Nightly evals and regression suites
- Bulk document summarisation and enrichment
- Embedding generation and index rebuilds
- Scheduled reports and digests
- Data labelling and synthetic data generation
- Backfills and migrations
- CI jobs that aren’t blocking a merge
Leave it — someone is waiting:
- Interactive chat and coding assistants
- User-facing API calls
- Anything with an SLA measured in seconds
Most organisations find 40-70% of total token volume is in the first category, and virtually all of it is scheduled by habit rather than requirement. A nightly job that runs at 09:00 UTC because someone picked that number in 2024 is paying double for nothing.
Batch APIs: The Same Trade, Different Shape
Providers without an hourly schedule offer the same economics through batch endpoints — you accept delayed completion, typically within 24 hours, for roughly 50% off. Anthropic’s Batch API, for example, discounts to $2.50/$12.50 on Opus-class models, and the discount stacks with prompt caching multipliers.
Functionally identical bargain: latency tolerance in exchange for money. The difference is who does the scheduling. With off-peak rates you control the clock; with batch APIs the provider does.
If your provider offers both, batch is usually the better tool — it needs no scheduler on your side and no timezone reasoning.
How To Implement It
1. Instrument first. Tag every API call with a deferrable boolean at the point it’s created. You cannot schedule what you can’t identify, and retrofitting this is much harder than adding it up front.
2. Add a queue, not a cron. A simple job queue that drains only during off-peak hours handles bursts, retries and backlogs correctly. A cron job firing at a fixed hour does not, and it will spill into peak when the backlog grows.
3. Set a deadline escape. Every deferred job should carry a maximum age. If a “cheap hours only” job is still queued after 18 hours, run it at peak. A saved dollar is not worth a missed report.
4. Alert on peak spend. Track the share of spend landing in peak windows. If it climbs above your expectation, something in the pipeline is firing at the wrong time — usually a retry storm.
5. Combine with caching. Off-peak halves the rate; prompt caching can cut input cost by 90%+ on top. They multiply.
The Risk Worth Naming
A provider that prices by the clock has learned something about its own leverage. Peak/off-peak arrived because demand outran capacity — DeepSeek’s schedule shipped alongside a general price increase of 51% to roughly 1,100%, not as a standalone discount.
Read it as a capacity signal. If your workload is genuinely latency-sensitive and can’t move, you’re now exposed to a provider whose peak rate is set by demand you don’t control. Keep a second provider warm and tested, and measure cost per completed task rather than per token, so that a schedule change doesn’t quietly reprice your product.
Last verified: August 17, 2026.