DeepSeek API Pricing Goes Live Today: V4-Pro GA and the Peak and Off-Peak Math

The vague warning is now a price list. On August 6, DeepSeek emailed API customers that it planned to raise prices “by a relatively large margin.” On August 13, it shipped the numbers alongside the DeepSeek-V4-Pro general availability release. And today, at 16:00 UTC, the new peak and off-peak pricing takes effect on every live API call.

For heavy AI users, this is the moment the cheap-AI assumptions break. The off-peak rate is billed as a discount off peak, but both rates are higher than what you were paying a week ago. If your pipelines treat DeepSeek as the bottom-of-the-cart pricing floor, this update rewrites that floor today. Here is exactly what changed, what it costs you, and how to route around it.

The new DeepSeek API pricing, confirmed

DeepSeek publishes four tiers of pricing per model: cache-hit input, cache-miss input, and output, each split into off-peak and peak. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC. Everything else is off-peak, and off-peak is exactly half the peak rate.

For deepseek-v4-flash, the new per-million-token prices are:

• Cache-hit input: $0.007 off-peak, $0.014 peak
• Cache-miss input: $0.22 off-peak, $0.44 peak
• Output: $0.66 off-peak, $1.32 peak

For the now-GA deepseek-v4-pro:

• Cache-hit input: $0.022 off-peak, $0.044 peak
• Cache-miss input: $0.66 off-peak, $1.32 peak
• Output: $1.98 off-peak, $3.96 peak

These replace the flat rates that were live until this morning: $0.14 and $0.28 per million for V4-Flash input and output, and $0.435 and $0.87 for V4-Pro.

The sticker shock is real, even off-peak

The framing of “off-peak is half of peak” is easy to misread as a discount. It is not a discount off anything you had access to. Compare like with like:

• V4-Flash output was $0.28. Off-peak output is now $0.66, about 2.4x higher. Peak is $1.32, almost 4.7x higher.
• V4-Pro output was $0.87. Off-peak is now $1.98, about 2.3x higher. Peak is $3.96, about 4.6x higher.
• Cache-miss input tokens rise roughly 1.5x off-peak and 3x at peak for both models.

The cheapest path on the platform, deepseek-v4-flash off-peak cache-miss input at $0.22 per million, is still very cheap by global standards. But anyone who built a cost model on the old $0.28 output rate just saw that line item multiply. If you route thousands of agentic output tokens a month through V4-Flash, the increase is not a rounding error.

Peak hours moved, and that changes your schedule

Earlier guidance put peak pricing around Beijing business hours, which conveniently fell outside North American and European working days. The final schedule is different. The announced peak windows, 01:00 to 04:00 and 06:00 to 10:00 UTC, overlap directly with the European morning and early working hours, and with the start of the North American East Coast workday.

For a developer in Paris or London, the 07:00 to 10:00 UTC window is peak for the first hours of the day. For a team on the US East Coast, 08:00 to 10:00 UTC is still early, but 06:00 to 10:00 covers the start of many automated and batch runs. Shifting work to the off-peak gaps, roughly 04:00 to 06:00 and 10:00 to 01:00 UTC, is where the 2x swing lives. The hours that used to be your safe off-peak default may now be the most expensive part of your day.

V4-Pro is now generally available

The pricing change rides on a bigger release. DeepSeek-V4-Pro moved from preview to general availability with model version DeepSeek-V4-Pro-0813, and it brings two changes that matter for cost control.

First, flexible reasoning effort. You can now set low, high, or max reasoning effort on both V4-Pro and V4-Flash. Low effort is for simple, high-volume tasks; high is the default for daily agent workflows; max is for complex multi-step problems. Reasoning effort is a direct lever on tokens consumed, so it is also a direct lever on how the new output prices hit you. Routing trivial lookups to low effort can hold spend down even after the rate increase.

Second, native OpenAI Responses API support, tuned for one-click Codex setup. That lowers the integration cost of pointing a Codex-style agent at DeepSeek, which matters for teams that were already considering it as a routing target. The Anthropic-format endpoint is also still available.

What this means for your routing strategy

The old playbook was simple: route the batchable, cheap-safe workload to V4-Flash off-peak and save. That playbook still works, but the savings floor moved up. The updated math to run:

• Batch and retry loops should run in the true off-peak gaps, not in the European or North American morning windows.
• Reach for low reasoning effort on anything that does not need deep multi-step reasoning. Output tokens are the most expensive line item now, and effort is the cheapest way to cut them.
• Cache discipline matters more. Cache-hit input at $0.007 to $0.014 per million is still a rounding error, while cache-miss input is 30 to 60x that. Keep shared prefixes cached to protect the input side of the bill.
• Re-baseline every cost comparison you made last month. DeepSeek is still cheap, but its gap to US providers narrowed, and the directional story changed from “prices are falling” to “prices are up.”

The bottom line for heavy users

The DeepSeek API pricing update is live today, and it is a real increase on every line item, even before you touch peak hours. V4-Pro is production-ready, flexible reasoning effort gives you a genuine dial on token burn, and the new UTC-based peak windows mean your old off-peak schedule may no longer be safe. Route flexible work into the genuine off-peak gaps, set effort to the minimum your task tolerates, and re-run your blast-radius numbers before your next batch goes out. The ultra-cheap DeepSeek price floor is gone. What replaces it is a provider that is still competitive, but only if you actually work the schedule.