DeepSeek V4 API Pricing Now Depends on the Hour

DeepSeek’s V4 API pricing changed from a flat schedule to peak and off-peak billing at 00:00 Beijing time on August 17, 2026—16:00 UTC on August 16. Off-peak rates are exactly half of peak rates, while the new prices range from 50% to 1,100% above the previous flat rates depending on the model, token type, and time of use. 12

What changed in DeepSeek V4 pricing?

The new schedule applies to DeepSeek-V4-Flash and DeepSeek-V4-Pro. Peak hours are:

• 09:00–12:00 Beijing time
• 14:00–18:00 Beijing time

All other hours are off-peak. A request is billed according to the period in which the platform receives it, so developers need to account for Beijing time when scheduling traffic. 27

The change is not simply a uniform price increase. It creates a 1:2 relationship between peak and off-peak rates: the same model call costs twice as much during either peak window.

DeepSeek V4-Flash and V4-Pro prices per million tokens

The following figures are in yuan per million tokens. “Cache hit” means the input can use a previously cached prompt prefix; “cache miss” means the input is processed at the higher uncached-input rate.

Model and token type	Previous flat rate	Off-peak rate	Change vs. previous	Peak rate	Change vs. previous
V4-Flash cache-hit input	¥0.02	¥0.05	+150%	¥0.10	+400%
V4-Flash cache-miss input	¥1.00	¥1.50	+50%	¥3.00	+200%
V4-Flash output	¥2.00	¥4.50	+125%	¥9.00	+350%
V4-Pro cache-hit input	¥0.025	¥0.15	+500%	¥0.30	+1,100%
V4-Pro cache-miss input	¥3.00	¥4.50	+50%	¥9.00	+200%
V4-Pro output	¥6.00	¥13.50	+125%	¥27.00	+350%

The Flash figures are documented as ¥0.05, ¥1.50, and ¥4.50 off-peak for cache-hit input, cache-miss input, and output, rising to ¥0.10, ¥3, and ¥9 during peak hours. 2 Reports on the broader V4 adjustment place the increases across models and token categories between 50% and 1,100%. 18

Which workloads are most affected?

Batch jobs can be moved to cheaper hours

Evaluation runs, document indexing, backfills, synthetic-data generation, and other non-urgent jobs can be queued outside the two peak windows. Because the off-peak rate is half the peak rate, scheduling alone can materially reduce the cost of flexible workloads. 2

This makes time zone a new operating variable for AI infrastructure. A queue that knows both the job’s deadline and the Beijing-time billing window can defer work without changing the model or reducing the token budget.

Interactive applications have less scheduling flexibility

Customer support, coding assistants, and real-time agents generally run when users are active. Those applications may have to absorb peak rates, but they can still reduce costs through prompt reuse, cache-friendly request design, shorter outputs, and model routing.

Caching is particularly important because cache-hit input remains far cheaper than cache-miss input on both models. For high-volume systems that repeatedly send the same instructions, tool definitions, or document context, improving cache-hit rates may matter more than small changes in request volume.

Output-heavy applications face a larger absolute bill

Long reasoning traces, generated reports, and large code responses consume output tokens. At peak, those tokens cost ¥9 per million for Flash and ¥27 per million for Pro. Developers can respond by setting tighter output limits, using Flash for routine turns, and reserving Pro for tasks where its additional capability justifies the premium. 12

Pro cache hits show the largest percentage increase

V4-Pro’s peak cache-hit price is 12 times its previous flat rate, the largest percentage increase in the new schedule. The absolute price remains small compared with uncached input and output, but the change is significant for applications that process very large volumes of repeatedly cached prompts. 119

How the August 13 V4-Pro release fits in

The pricing change followed the general availability of DeepSeek-V4-Pro-0813 behind the deepseek-v4-pro endpoint in August. Reported specifications describe a text model with a one-million-token context window, a maximum output of 384,000 tokens, and a mixture-of-experts architecture with approximately 1.6 trillion total parameters and 49 billion active parameters per token. 2933

DeepSeek’s documented account-level concurrency limits also distinguish the two tiers:

• V4-Pro: 500 simultaneous requests
• V4-Flash: 2,500 simultaneous requests

A request occupies a slot through completion, including streamed responses. 28

Taken together, the pricing and concurrency structures suggest a product segmentation in which Flash is better suited to high-throughput routine traffic, while Pro is positioned as a more capacity-constrained tier for demanding reasoning and large-context workloads. That positioning is an inference from the published economics and limits, not a direct statement about every use case. 228

What time-of-use pricing signals about AI infrastructure

DeepSeek’s move resembles demand shaping for GPU inference. Providers face concentrated demand during the hours when interactive users are online, while batch workloads can often wait. Charging more during those periods and less at other times creates an incentive to spread computation across the day.

For developers, that changes the optimization problem. Cost control is no longer only about choosing a model and counting tokens. Production teams must also decide:

1. Which requests are latency-critical.
2. Which jobs can enter an off-peak queue.
3. How much prompt content can be reused through caching.
4. When to route a task to Flash instead of Pro.
5. How much generated output the application actually needs.

The broader implication is a shift away from flat-rate price competition toward pricing based on capacity scarcity. DeepSeek is still offering a lower-cost lane for flexible demand, but it is also making the cost of high-end inference more sensitive to when that capacity is used. Reuters described the change as a substantial increase that varies by model, token category, and usage period—not as a single across-the-board surcharge. 1

For teams using DeepSeek V4, the practical lesson is straightforward: treat scheduling, caching, routing, and output controls as part of the API cost model. A workload that cannot be moved in time may need a different model or tighter token budget; a workload that can wait may be able to preserve much of its economics by running off-peak.