Keep your weekend AI build near zero with off-peak batches

Your weekend AI build can stay near zero if you decide when each request is allowed to happen. The API bill is a calendar, not a rate card. Move the heavy work into the quiet hours, and the project stays cheap even when the price jumps.

DeepSeek moved its API to a peak/off-peak schedule at 16:00 UTC on August 16, 2026, with peak hours at 01:00–04:00 and 06:00–10:00 UTC and off-peak priced at half peak. V4-Pro output tokens changed from a flat $0.87 per million to $1.98 off-peak and $3.96 at peak. V4-Flash output tokens changed from a flat $0.28 per million to $0.66 off-peak and $1.32 at peak.

DeepSeek said the revision was intended to allocate resources more reasonably and move developer workloads toward less congested periods. For a small side project, the model can stay the same; the bill changes because the same tokens now arrive at different times.

The weekend API budget rule: classify workloads as interactive, batch, or peak-sensitive; schedule batch off-peak; set a peak-hour fallback model and a spend cap.

The schedule is the budget

Classify the work before you touch the code. A weekend project usually has a person waiting for a reply, a large job that can finish while you sleep, and a heavy call that would be expensive if it lands in a busy window.

Interactive work is anything where a user is waiting: chat, search, form help, quick summaries. Batch work is anything that can finish overnight: dataset labeling, long rewrites, report generation, backfilling old content. Peak-sensitive work is heavy generation that you would normally run during the day but can move or reroute.

Use a simple test: keep a waiting human on the cheap interactive path. Mark queued output as batch. Mark a long, expensive job that is likely to run while you are awake as peak-sensitive. You do not need perfect labels. You need a default for each job. Keep a short list in your notes: what runs now, what runs later, and what needs a fallback.

The batch runs in the cheap hours

Schedule the batch into the off-peak window, not the window that feels quiet to you. Check the conversion before you celebrate a cheap overnight run, because your local evening may not match UTC. Under official concurrency limits, V4-Flash supports 2,500 simultaneous requests and V4-Pro supports 500. V4-Flash is the easier model for parallel batch work. A log that shows the start time is enough to prove the job ran in the cheap window.

Start with a single batch job, not the whole app. Pick the heaviest thing you already run: a nightly digest, a content refresh, a data cleanup. Give it a start time, a timeout, and a place to write results. Let it fail quietly and retry later, not waste your weekend budget.

The fallback handles the busy hours

Set the fallback before the peak hits. GPT-5.6 Luna's published output price is $1.20 per million tokens, making V4-Flash cheaper off-peak but slightly more expensive at peak on output alone. During Beijing business hours, V4-Pro's peak output price exceeds GPT-5.6 Luna's output price. A fallback helps heavy jobs that cannot wait.

Add a spend cap: a daily limit, an alert, and a queue that holds the rest. Keep the fallback boring: a single model, a single route, a single cap. Use the peak-hour model when it is cheaper for that job; queue the job when it is not. Keep the cap low enough that a bad loop cannot surprise you, and high enough that a normal weekend does not stall. A simple alert is enough: a message when the day is almost spent, and a pause when it is spent. Then you can close the laptop and let the queue finish.