DeepSeek-V4-Pro Goes GA With Peak and Off-Peak API Rates

DeepSeek put DeepSeek-V4-Pro into general availability on its app and web Expert Mode and on the API, and it attached a peak and off-peak price schedule that takes effect at 16:00 UTC on 16 August 2026 (23:00 WIB). Off-peak V4-Pro output is $1.98 per million tokens.

What DeepSeek-V4-Pro ships at GA

The model is live for people who pick Expert Mode in the DeepSeek app or on the web, and it is live for developers who already call the API. Teams that hard-coded an older DeepSeek route still need to treat this as a new product and a new bill, not as a silent patch.

Reasoning effort is no longer a single hidden default. DeepSeek-V4-Pro accepts three effort settings: low, high, and max. Low is the light pass. High is what DeepSeek calls the everyday agent setting. Max is the heavy pass. That three-way split is the kind of control a skill prompt should expose as a field, not bury in a paragraph of instructions.

The API also speaks the OpenAI Responses API natively, with a one-click Codex setup. If your wrapper already talks Responses, the integration path is shorter than a custom client. If you route through a gateway, treat DeepSeek-V4-Pro as one more named route with its own effort flag and its own clock.

Peak and off-peak rates change the bill

The pricing story is the one operators will feel first. DeepSeek is leaving a flat rate behind. New prices start at 16:00 UTC on 16 August 2026. Peak windows are 01:00–04:00 UTC (08:00–11:00 WIB) and 06:00–10:00 UTC (13:00–17:00 WIB). Off-peak is 50 percent of peak, so the remaining hours in the day sit at the lower band.

For DeepSeek-V4-Pro output, off-peak is $1.98 per million tokens and peak is $3.96 per million. The prior output price was a flat $0.87 per million. That is a jump even if every token lands off-peak, and a larger jump if interactive traffic sits inside the two peak windows.

Do not invent a blended rate. The official schedule is time of day plus the output rate. Interactive chat that must answer now will pay peak when the clock says peak. Batch evals, overnight document passes, and other deferrable jobs can wait for off-peak. That is the same scheduling habit cloud teams already use for compute. It is now an inference habit too.

A skill that always calls DeepSeek-V4-Pro at max effort the moment a ticket arrives will eat the new peak output price. A skill that sets effort from the job type, and that queues non-urgent work outside 01:00–04:00 and 06:00–10:00 UTC, will not.

Why the clock belongs in the skill, not the prompt

Put the model, the effort, and the schedule in the wrapper. The prompt should name the job. The route should name DeepSeek-V4-Pro, low or high or max, and whether the run can wait. If a run cannot wait, log that it paid peak. If it can wait, do not pretend a 02:00 UTC call was a bargain.

Single-vendor lock-in is still a risk even when the vendor used to be cheap. Hosted forges go down. Labs change prices. Gateways exist because teams got tired of rewriting clients. DeepSeek-V4-Pro being generally available does not change that. It gives you a stronger model and a more complicated invoice.

Keep a fallback route. Keep cache-aware prompts if your stack already prefixes shared instructions. DeepSeek did not publish a cache figure in the facts used here, so do not invent a hit-rate savings. Just stop sending a fresh wall of system text on every turn if you do not have to.

If you publish this update in a changelog, the citation sentence is DeepSeek-V4-Pro, 16 August 2026, 16:00 UTC, $1.98 off-peak and $3.96 peak output. Off-peak is half of peak. The old flat output price was $0.87. Those are the numbers. Everything else is scheduling.

Sources

• DeepSeek-V4-Pro GA Release — DeepSeek API Docs, 16 August 2026
• DeepSeek Raises API Prices, Putting Developers' AI Economics to the Test — ADTmag, 20 August 2026