DeepSeek V4 Pricing Rewards Workloads That Can Wait
Matthew Leo · Published August 14, 2026 · Last updated September 29, 2026 · AI
DeepSeek has turned the time of day into part of the price of running an AI model.
Its V4 Flash and V4 Pro APIs will cost twice as much during two daily peak windows as they do during the rest of the day. That gives developers a meaningful discount for batch work that can wait, but it also makes a simple price-per-million-tokens comparison incomplete.
For Canadian teams, the sensible calculation includes the workload schedule, US-dollar billing, data-handling terms and the cost of being unavailable when a cheaper service is congested.
The same model has two price cards
DeepSeek’s official pricing documentation says the new rates take effect at 16:00 UTC on August 16, 2026. Peak periods run from 01:00 to 04:00 UTC and 06:00 to 10:00 UTC. All other hours receive the off-peak rate.
| One million tokens | Flash off-peak | Flash peak | Pro off-peak | Pro peak |
|---|---|---|---|---|
| Cache hit input | US$0.007 | US$0.014 | US$0.022 | US$0.044 |
| Cache miss input | US$0.22 | US$0.44 | US$0.66 | US$1.32 |
| Output | US$0.66 | US$1.32 | US$1.98 | US$3.96 |
The discount is exactly half, but the relevant clock is UTC. In Toronto during daylight time, the listed peak windows correspond to 9 p.m. to midnight and 2 a.m. to 6 a.m. the previous local calendar day. Teams elsewhere in Canada must convert for their own time zone and daylight-saving period.
Reuters reported the V4 Pro release on August 13 and highlighted its substantially higher price than V4 Flash. Those ratios are accurate within a given pricing window, but they do not tell a team what one real job will cost.
A sample workload makes the difference clearer
Consider a monthly batch process using 100 million cache-miss input tokens and 20 million output tokens. It might summarize internal documentation, classify support requests or generate draft descriptions that are reviewed later.
| Model and window | Input cost | Output cost | Total |
|---|---|---|---|
| V4 Flash off-peak | US$22.00 | US$13.20 | US$35.20 |
| V4 Flash peak | US$44.00 | US$26.40 | US$70.40 |
| V4 Pro off-peak | US$66.00 | US$39.60 | US$105.60 |
| V4 Pro peak | US$132.00 | US$79.20 | US$211.20 |
The arithmetic is simple: multiply token volumes by the published per-million rate. The operational conclusion matters more. Moving an asynchronous V4 Pro job outside peak hours saves US$105.60 in this example without changing the model.
That does not mean every request should wait. Customer support, interactive coding and fraud review may be valuable precisely because they respond immediately. Scheduling a delay that costs staff time or loses a customer would be a false economy.
Caching needs engineering, not wishful budgeting
The cache-hit rate is dramatically lower than the cache-miss rate. A team should not build its forecast by assuming every repeated prompt will qualify.
Measure actual cache hits in production and separate stable context from material that changes on every request. If a large system prompt or reference document can be reused, caching may matter more than the time window. If each request contains unique documents, budget closer to the cache-miss rate.
Also track output length. Generative applications often spend more on output than expected because a prompt that asks for “a complete analysis” has no practical stopping rule. Token limits and concise output formats are cost controls, not merely interface choices.
The cheapest token can carry a larger organizational cost
DeepSeek’s table is denominated in US dollars. Canadian buyers must account for exchange rates, card or payment fees and any applicable taxes. Those items can move a small proof-of-concept budget, though they do not erase the underlying discount.
The larger question is whether the service is appropriate for the data. Before sending proprietary code, personal information or confidential documents, review the current terms, privacy documentation, retention rules and processing locations that apply to the specific account and API. Do not assume an API automatically receives the same controls as an enterprise contract.
A useful procurement record should identify the model ID, pricing date, region, approved data classes and the person who reviewed the terms. A low inference price does not grant a developer permission to submit every repository.
Run a two-week cost and quality test
Compare Flash and Pro on a representative set of tasks, not vendor examples. Record input, cached input, output, latency, failures and the amount of human correction each result required.
Then schedule non-urgent work into the off-peak window and measure whether latency or reliability changes. A model that needs twice as much review can remain more expensive even when its token bill is half the price.
DeepSeek’s time-based discount is genuine and unusually legible. It rewards queues, nightly processing and other work that can move. The decision for Canadian developers is not whether off-peak is cheaper. It is which work can safely wait, which data can leave the organization and whether the measured result is worth more than the number on the price card.
Tags: DeepSeek, AI pricing, developers, cloud costs, Canada