DeepSeek API prices rise August 16 — output up as much as 371%

Five days ago this page wrote about Gemini 3.7 Flash's price being cut in half. The day before that, DeepSeek's V4-Pro launch.

Two consecutive entries recording AI model prices going down. On August 16 the opposite takes effect.

DeepSeek's API prices rise. Output tokens by as much as 371%.

1. The numbers

V4-Flash (per million tokens, USD)

V4-Pro (per million tokens, USD)

The largest increase is not in either table. It is on cache-hit input tokens, which rise between 52% and roughly 1,100% depending on the line item. Cache hits are the discount path for repeated prompts — precisely what conversational products and looping agents lean on hardest. The cheapest line on the sheet took the biggest increase.

2. When peak falls, depending on where you are

The windows are fixed in UTC, so who pays the premium depends on time zone. For a team in Seoul, peak covers most of the working day — 10am to 1pm and 3pm to 7pm. For a team on the US east coast, both peak windows fall largely outside business hours, which makes the same price change materially cheaper.

Off-peak is exactly half of peak. The change starts at 16:00 UTC on August 16, 2026.

3. DeepSeek's reason, and the other reading

The company frames it as allocation: the goal is to "allocate resources more reasonably" and to have developers "schedule their tasks based on actual usage." Push batch work into quiet hours and it runs at half price; the intent is to flatten the load curve.

An analyst's summary is shorter: demand rose, supply is constrained, so price rose.

There is also a timing problem. In May 2026 DeepSeek said it would make its discounted promotional pricing permanent. This announcement reverses that within three months.

4. Three companies, one week, opposite directions

Two lowered the cost of entry; one raised it.

Reading that as DeepSeek losing the competition would be premature. The company's account is that demand is outrunning capacity — and if that is accurate, a price rise is a sign of strength, not weakness. This article has no basis for deciding which reading is right. Usage data after the change is what settles it.

5. What is not settled

The official price table was not checked directly. The cache-hit figures in particular are press tallies.

The time-zone basis is unconfirmed. Whether peak windows are a single UTC standard for all customers, regardless of location, was not established.

Grace periods are unknown. Whether existing contract customers are exempted has not been confirmed.

The underlying structure of token pricing — why output costs more than input, and what cache hits are — is in "What AI token prices are." The new time-of-day axis is covered in "What peak pricing for AI APIs is."