How off-peak pricing works
Most LLM APIs charge the same per-token rate around the clock. An off-peak scheme splits the day instead: the vendor names the hours when its capacity is busiest and charges the full rate only then. DeepSeek publishes its peak hours as 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday. Every other hour is off-peak, including weekends, and its pricing page adds that Chinese public holidays count as off-peak in full.
Off-peak rates are half the peak rates on DeepSeek's page, and the cut applies to every meter: cache-miss input, cache-hit input and output. DeepSeek V4.1 Flash, for example, lists $0.30 per million input tokens at peak and $0.15 off-peak; output goes from $1.20 to $0.60. Nothing in the request changes — the rate is set by when the request is processed.
In our indexOff-peak rates in our index
Counted at build from our price index: published off-peak rates against the same model's peak rates among the 167 current public models we price, from 33 vendors. Each rate comes from the vendor's own price page, read on 2026-09-19. How we collect prices.
Planning work around the window
The peak hours fall inside the Chinese business day and the European morning. In UTC terms, the cheap hours on weekdays are 04:00–06:00 and 10:00–01:00 — which covers the whole American working day. A team in the Americas running interactive traffic is therefore already paying off-peak rates for most of its usage; a team in Asia gets the discount by moving deferrable jobs to its evening and weekends.
Deferrable work is the target: nightly enrichment, re-scoring a document store, evaluation runs, synthetic data. Schedule the job to start after 10:00 UTC on a weekday or any time at the weekend, and make it pause rather than continue into the next peak window if it runs long.
Off-peak pricing vs a batch discount
Both trade timing for price, but differently. A batch API discounts any request you are willing to wait up to 24 hours for, at any time of day, through a separate endpoint. Off-peak pricing discounts real-time requests on the normal endpoint, but only inside a fixed clock window. Off-peak suits work that needs answers immediately but can choose when to start; batch suits work that can wait for its answers but needs to start now.
Caching stacks with the off-peak rate on DeepSeek, because cache-hit input has its own off-peak price. A long shared prefix run at the weekend gets both discounts.
Watching for changes
A clock-based discount is a pricing policy, and DeepSeek's page notes it may adjust prices and recommends checking regularly. Our index stores the peak rate as the model's list price and the off-peak rate beside it, both dated to the day we read DeepSeek's page, so a change to the window or the size of the cut shows up as a dated change on the DeepSeek pricing hub.