What Is Off-Peak API Pricing? Time-of-Day Token Costs Explained

Off-peak API pricing charges you less for the same tokens simply because of when you send them. It's borrowed straight from electricity markets — and it's starting to appear in LLM APIs. Here's how it works and how to use it.

For years, LLM APIs had one price per model, all day, every day. That's changing. Some providers now split the day into peak hours (high demand, higher price) and off-peak hours (low demand, lower price). The tokens are identical; only the clock differs.

Why Providers Do This

GPU capacity is fixed in the short term, but demand is not. During busy hours, a provider's clusters run hot and requests queue. During quiet hours, expensive hardware sits idle. Time-of-day pricing is a lever to smooth that curve: make off-peak cheaper and price-sensitive batch workloads shift into the valleys, freeing peak capacity for latency-sensitive, willing-to-pay traffic.

It's the same logic as off-peak electricity tariffs or cloud spot instances. The provider trades a lower price for demand it can schedule around; you trade timing flexibility for a discount.

How the Billing Works

The typical structure is a defined peak window in UTC, with a fixed multiplier. DeepSeek, for example, moved its V4 Flash and Pro models to this model in August 2026: peak hours of 01:00–04:00 and 06:00–10:00 UTC, with peak rates exactly 2x the off-peak base. A request is billed at whatever tier is active when it's processed.

Key details to check for any provider that offers this:

• The window, in your timezone. Peak is defined in UTC. Convert it to local time so you know when your own work lands.
• The multiplier. Is peak 1.5x, 2x, or more? That sets the size of the prize for shifting.
• Whether it applies to input, output, or both. Usually both, but confirm.
• Which tier is the "sticker price." Providers often quote the off-peak rate as the headline — but you only get it outside peak hours.

How to Save With It

The strategy is simple: move time-flexible work off-peak, and let interactive work fall where it may.

• Schedule batch jobs. Nightly code review, test generation, documentation passes, embeddings backfills, and bulk refactors have no human waiting — queue them for off-peak windows with a cron job or a job scheduler.
• Don't micro-optimize interactive work. A developer waiting on a live completion shouldn't be blocked to save a few cents. Pay peak for latency-sensitive tasks.
• Check your natural overlap. If your team's working hours already sit mostly outside the peak UTC window, you may be getting off-peak rates without doing anything.

Off-Peak vs Batch API — Not the Same Thing

Don't confuse the two. Batch API pricing gives a discount (often 50%) for asynchronous jobs the provider can process within a window (typically 24 hours), regardless of clock time. Off-peak pricing gives a discount for synchronous requests that happen to run during low-demand hours. Some providers offer both, and they can stack: a batch job scheduled to run off-peak captures both discounts.

Is This the Future?

Probably, for capacity-constrained providers. As frontier demand keeps outpacing GPU supply, exposing time-of-day signals to customers is an efficient way to ration. The practical implication for you: the single-number price tag is fragmenting into conditional rates, and "what does this model cost?" increasingly has the answer "it depends when you run it." Reading the schedule is now part of cost management.

Compare model rates and estimate your project's token spend in our AI Cost Calculator.