DeepSeek Raised Prices Today. Here's the Real Bill.
The new peak/off-peak API pricing went live on August 17. Which workloads actually got more expensive, which barely moved, and three ways to keep your bill low.
DeepSeek's new API pricing went live at 00:00 Beijing time today, August 17, and V4-Pro became a fully commercial service on the same day. The headlines are real: on the most extreme line item, the price rose about 12 times. But most headlines stop there, and the details matter more. Some workloads barely moved, a few are still nearly free, and the structure of the new pricing tells you exactly what DeepSeek wants you to do: move your jobs out of peak hours.
The new price table
DeepSeek split every rate into peak and off-peak. Peak hours are 01:00-04:00 and 06:00-10:00 UTC, which is 09:00-12:00 and 14:00-18:00 in Beijing. Off-peak rates are exactly half of peak rates, so the table below (official USD list prices per 1M tokens) is all you need:
| Model | Input, cache hit | Input, cache miss | Output |
|---|---|---|---|
| V4-Flash off-peak | $0.007 | $0.22 | $0.66 |
| V4-Flash peak | $0.014 | $0.44 | $1.32 |
| V4-Pro off-peak | $0.022 | $0.66 | $1.98 |
| V4-Pro peak | $0.044 | $1.32 | $3.96 |
In yuan, the numbers most Chinese developers will see are the off-peak ones: V4-Flash at ¥0.05 / ¥1.5 / ¥4.5 and V4-Pro at ¥0.15 / ¥4.5 / ¥13.5, all doubled during peak hours.
What the "11x" headline is really about
The largest jump is real but narrow. V4-Pro's cache-hit input rate went from ¥0.025 per million tokens to ¥0.30 during peak hours, about 12 times the old price. That is where the "prices up 11x" and "up to 1100%" headlines come from.
Output did not get that treatment. V4-Pro output went from ¥6 to ¥27 at peak, 4.5x, and to ¥13.5 off-peak, about 2.25x the old rate. Off-peak rates on the same models were never part of the old pricing, so they are new baselines rather than hikes on top of hikes. The honest summary: the biggest increases hit the line items that were already almost free, and they only apply during six hours of the day.
What this means for your bill
Against Western frontier models, DeepSeek still looks like a discount, just a smaller one. V4-Pro output at peak is $3.96 per million tokens, compared with roughly $50 for Claude Fable 5 output, about 13x cheaper. Off-peak, it is $1.98, about 25x cheaper. V4-Flash output at peak is $1.32, about 38x cheaper. The "1/10th of the price" era is over for peak usage, but the gap is still a wide one.
The real story is cache hits. Cache-hit input costs $0.007 to $0.044 per million tokens, depending on model and time. A million tokens of cached system prompt costs about the same as a few seconds of electricity. For workloads with stable context, the bill barely changed.
Three ways to keep costs down
- Shift jobs off the six peak hours. This is the pricing's explicit purpose. DeepSeek says it introduced peak/off-peak rates to balance compute load and push developers to schedule off-peak. Batch jobs, nightly syncs, and re-indexing runs cost half as much after 18:00 Beijing time.
- Structure prompts so they hit the cache. Reuse a stable system prompt, keep the codebase context in a fixed order, and let the cache do its work. At $0.007 per million tokens for a cache hit, your static input is effectively free.
- Use the right model for the job. V4-Flash at $0.44 cache-miss input and $1.32 output during peak is still cheaper than most competitors' off-peak rates. Reserve V4-Pro for the tasks that need the stronger model.
The bottom line
The price increase is real and it lands mostly on peak-hour usage of already-expensive line items. For off-peak, cache-friendly workloads, DeepSeek is still the cheapest frontier-class option by a wide margin. We updated our DeepSeek card with the confirmed rates and the August 17 verification date, and the card will tell you the day we last checked the numbers.