Off-peak AI API pricing: have software coded overnight for half the price

An AI provider introduced off-peak rates at half daytime pricing. The idea turns a budget question into an organisational one: which project tasks can wait overnight, and which block you until complete?

Off-peak pricing has arrived in AI APIs, the usage-billed channels through which software calls models. DeepSeek changed pricing on August 16, 2026 at 16:00 UTC, 6 p.m. in Paris, adopting peak and off-peak rates 'with off-peak prices set at half of the peak-hour prices' (DeepSeek API changelog). Off-peak costs half the peak rate.

What off-peak pricing means here

The principle resembles electricity night rates: identical consumption costs less when machines are idle. For a project owner it is easy to formulate and awkward to apply. Long jobs needing no presence, data conversion, whole-product verification, existing-software takeover, can start in the evening and be read at breakfast. Short work involving decisions each round, scoping, specification review, a correction you want immediately, gains nothing by waiting overnight.

What DeepSeek's page does not say

We read the source rather than its retelling. It gives the effective date, time, and half-price principle. It gives neither exact off-peak windows nor per-model amounts. An article claiming '8 p.m. to 8 a.m.' invents precision the publisher did not provide. Retain the scale and verify the window on your provider's pricing page on the day you decide to reschedule anything.

The other reduction: price per unit of work

Hourly prices move, as do work-unit prices. Cognition, Devin's publisher, released a model on July 8, 2026 presented as 'Frontier Intelligence at a Fraction of the Cost', then on July 13 an architecture intended to lower premium-model usage costs, noting 'Fable 5 costs twice as much per token as Opus 4.8' while working better (Cognition blog index consulted September 2, 2026; posts not individually opened). A pricier unit can therefore yield a cheaper completed task, and a reduction can hide in an unread line. A six-month-old profitability calculation is outdated, and tokens fall while bills rise for the same reason: call numbers explode, not unit prices.

What Maestro does and does not do

We do not control DeepSeek: connected assistants are Claude Code, Codex, Cursor, and Mistral Vibe, switchable mid-conversation. We do not schedule by clock either; no button says 'launch at 2 a.m.'. We do have the mechanics making delayed spending manageable: a queue preventing parallel consumption, a three-times-estimate cap before development, and a monthly budget across projects. A number visible before spending beats discovering a night rate afterward.

What matters more than the hour

Moving work overnight halves an amount at best with a particular provider. Rebuilding a feature three times because nobody wrote its intended behaviour triples it at any hour. In our campaigns, looping repairs moved the bill, since capped at two attempts per project before returning decisions to a human. Night pricing provides a discount; scoping removes the expense.

What you can do this week

Open your assistant provider's pricing page and check whether off-peak rates exist and their local hours. Then list current tasks needing no involvement between launch and result. If the list is short, night rates will not change your budget, and time is better spent setting a spending cap than watching a clock.