DeepSeek Ships V4-Pro and Changes API Pricing on August 16

DeepSeek made DeepSeek-V4-Pro generally available on August 13 ¹. Announced alongside it was a change to API pricing: the new schedule takes effect at 16:00 UTC on August 16, 2026 ¹².

The change has two parts. Unit prices go up, and billing moves to peak and off-peak rates that vary by time of day ². DeepSeek says the goal is “to allocate resources more reasonably,” setting off-peak prices at half the peak-hour prices to encourage users to schedule their tasks based on actual usage ³.

The Increase Ranges From Roughly 1.5x to 12x Depending on the Line Item

DeepSeek’s pricing page lists the current and new prices side by side ². In US dollars per 1M tokens, current pricing is:

Model	Input (cache hit)	Input (cache miss)	Output
deepseek-v4-flash	$0.0028	$0.14	$0.28
deepseek-v4-pro	$0.003625	$0.435	$0.87

From 16:00 UTC on August 16 ²:

Model	Tier	Input (cache hit)	Input (cache miss)	Output
deepseek-v4-flash	Off-peak	$0.007	$0.22	$0.66
deepseek-v4-flash	Peak	$0.014	$0.44	$1.32
deepseek-v4-pro	Off-peak	$0.022	$0.66	$1.98
deepseek-v4-pro	Peak	$0.044	$1.32	$3.96

Peak hours are two blocks, 01:00-04:00 and 06:00-10:00 UTC; all other hours are off-peak ².

Working from the table, the size of the increase varies quite a bit by line item. The deepseek-v4-pro output rate goes from $0.87 to $3.96 at peak, roughly 4.55x, and to $1.98 off-peak, roughly 2.28x. Cache-miss input goes from $0.435 to $1.32 at peak, roughly 3.03x. The largest multiple is on cache-hit input: $0.003625 to $0.044 at peak is roughly 12.1x. deepseek-v4-flash follows the same pattern, with output about 4.71x at peak and cache-hit input exactly 5x.

That cache-hit multiple matters most if you repeatedly send long conversation histories or codebases. For the same total number of tokens, a setup where a larger share hits the cache takes a proportionally bigger increase under the new schedule. DeepSeek does not state any multiples in its own materials; the figures above are calculated from the published price tables.

V4-Pro Runs Under “Expert Mode,” With Three Levels of Thinking Effort

DeepSeek-V4-Pro, generally available as of the same day, is reachable in the app and on the web through “Expert Mode” ¹. On the API side the model name is unchanged — specifying deepseek-v4-pro gets the latest version ¹³. The model table on the pricing page lists the version as DeepSeek-V4-Pro-0813 ².

DeepSeek says the GA version improves agent capabilities, with particularly significant gains in production environments ³. The benchmark figures it publishes are Terminal Bench 2.1 at 87.9, NL2Repo at 61.5, DeepSWE at 62.7, Toolathlon-Verified at 74.1, Cybergym at 83.3, Agents’ Last Exam at 25.7, and HLE at 42.7 without tools and 60.0 with tools ³. DSBench-FullStack at 71.1 and DSBench-Hard at 67.2 come from internal test sets DeepSeek maintains for full-stack development and hard coding-agent problems ³. All of these are DeepSeek’s own measurements, not third-party verification.

Two API changes come with the release. First, thinking effort on both V4-Pro and V4-Flash now has three levels: low, high, and max ³. DeepSeek suggests low for simple tasks, high for daily agent workflows, and max for more complex scenarios ¹³. Second, the API natively supports the OpenAI Responses API format, adapted for Codex, with a one-click configuration script ³.

Both models have a 1M-token context window and a maximum output of 384K tokens, with concurrency limits of 2500 for deepseek-v4-flash and 500 for deepseek-v4-pro ².

Moving Against the Direction of the Past Month

For about a month, headline API prices had been moving down. On July 30, OpenAI cut GPT-5.6 Luna by 80% and Terra by 20%. On August 13, an Ultrafast tier that raises speed while keeping the same model was announced as a limited preview (no pricing has been published for that tier). Rates coming down, or service split into tiers — and now DeepSeek has taken a third route: raise the base rate and discount by time of day.

DeepSeek’s own sequence leading here is visible in its changelog. On April 24 the API added V4-Pro and V4-Flash, and the legacy model names deepseek-chat and deepseek-reasoner were announced for retirement three months later, on July 24 ³. On July 31, V4-Flash reached its official release, and that entry said the official release of DeepSeek-V4-Pro would follow soon ³. The V4-Pro GA lands as announced, and the pricing change is timed to the point where the V4 lineup is complete ¹.

In practice, the time-of-day billing may require more thought than the unit increase itself. Interactive use cannot easily be moved, but batch jobs, scheduled runs, and agent workloads that run unattended cost half as much off-peak. Conversely, any automation running without regard to the clock may see a different bill for the same volume of work. Since the effective moment is precise — 16:00 UTC on August 16 — it is worth mapping which jobs run in which window before then.

DeepSeek also notes on the pricing page that “product prices may vary and DeepSeek reserves the right to adjust them,” asking users to check the page regularly for current pricing ². A side-by-side of vendor pricing plans is collected in our AI coding tool pricing comparison (as of July 11, 2026).