DeepSeek's Top Model Gets 77% Cheaper Tomorrow, and Its Peak Hours Are Your Working Morning

DeepSeek Is Routing Its Top Model to a Cheaper One, and Billing You at the Cheaper Price

DeepSeek has notified its API users that V4.1 Flash ships on 10 September 2026. The news is not the launch. It is what follows it: from 4:00 UTC, 7:00 in Riyadh, every request aimed at the flagship V4 Pro will be routed to the new model and billed at the new model's price. Anyone using the strongest model today pays the fastest model's price tomorrow, without changing a line of their system.

That is up to 77% off input and 70% off output. The number matters, but three details in the announcement matter more, and they are the ones that actually decide your bill.

The Numbers

Prices are per one million tokens, a token being the smallest unit of text the model reads: a word, a number, a punctuation mark. These are off-peak rates; peak rates are exactly double.

Item	V4 Pro today	V4.1 Flash tomorrow	Change
1M input tokens, read from cache	$0.022	$0.003	-86%
1M input tokens, fresh	$0.66	$0.15	-77%
1M output tokens	$1.98	$0.60	-70%

If you are on the regular V4 Flash today the drop is smaller but real: fresh input from $0.22 to $0.15, and cached input from $0.007 to $0.003.

Their Peak Hours Sit Inside Your Working Morning

DeepSeek bills at two rates, peak and off-peak, and the gap is double. Peak hours are published in UTC: 01:00 to 04:00 and 06:00 to 10:00, Monday to Friday.

Convert that to Riyadh time and peak runs from 04:00 to 07:00 and from 09:00 to 13:00. The entire Saudi working morning falls inside peak, at twice the price. After 13:00 the rate halves until the next morning.

The practical consequence is direct. Work that answers a customer in real time stays where it is, whatever it costs, because the customer does not wait. Work that nobody is waiting on — monthly report generation, classifying a message archive, extracting data from old files, summarising minutes — costs half as much if it runs after 13:00 or overnight, with no other change. Scheduling is an operational decision, not a technical one, and here it moves the bill as much as changing the model does.

The Cache: A Fifty-Fold Difference

Look at the first row of the table again. Fresh input is $0.15. Input read from cache is $0.003. That is a fifty-fold gap, the largest single spread in the whole price list, and the most commonly ignored.

The cache works on the beginning of your message. Send the same opening text on every request and DeepSeek charges it once at full price, then reads it back for a fraction. Most real systems carry a large fixed block on every call: the assistant's instructions, the product and price list, the returns policy, the tone of voice. None of that changes between one customer and the next.

So the difference between a system that orders its prompt and one that does not is not an engineering nicety. It is whether the fixed part is charged once or recharged in full on every message. The rule is simple: fixed first, variable last.

Take an example. A WhatsApp assistant handling 3,000 conversations a month, each with 4,000 input tokens and 500 output tokens, costs about $10.89 a month at today's off-peak V4 Pro price. At tomorrow's V4.1 Flash price it falls to $2.70. Order the prompt so 80% of the input reads from cache and it falls to $1.29. Eight times between the first figure and the last, and in the final step the model never changed. Only the order of the text did.

The Clause Worth Reading Rather Than Skipping

One sentence in the notice deserves attention: continued use of the service after the billing change is treated as acceptance of the new terms, and anyone who disagrees may cancel and request a refund.

This is standard language in cloud service contracts, and it is unobjectionable here because the change favours the user. But the principle is worth noticing: the same clause permits a price change in the other direction by the same mechanism. If you are building a commercial system on top of somebody else's API, your running cost is not a fixed number in the business case. It is a term that changes on notice.

The Lesson Beyond DeepSeek

Notice what happened to the model name. A developer who wrote deepseek-v4-pro into their system will find their requests going somewhere else by the provider's decision. Here the outcome favours them: cheaper, faster, better by the vendor's own account. But the same mechanism is what moves your requests to a different model on a day when the decision does not favour you.

One rule covers it: do not scatter the model name through your system. Keep it as a single configuration value, and let every call pass through one layer in your code. Then changing the model, or changing the provider entirely, is a one-line edit rather than a project. It costs nothing to do at the start and weeks to retrofit later.

And the point that matters most to a business owner: at these prices the model is no longer what stops a project. Running an assistant across thousands of conversations a month now costs less than your office coffee bill. What decides the outcome is everything around the model: whether your data is organised enough for it to read, whether it is actually wired into your systems or answering from nothing, and whether you know what happens when it gets something wrong.

How We Handle This at Origami

In the systems we build for clients, every model call passes through a single layer, and the model name lives in a configuration file rather than repeated through the code. When the market shifts, and it shifts about monthly, moving a client to a cheaper or more accurate model is a decision executed in minutes and measured afterwards, not a rebuild.

We also treat prompt order as a design decision rather than a detail: the fixed block first so it reads from cache, the variable part last. And we separate what needs an immediate answer from what can wait, so the second kind runs in the cheap hours. Three decisions nobody sees in the interface, all of which show up on the invoice.

Sources

• DeepSeek — official Models & Pricing page (V4 Pro and V4 Flash rates, peak hours)
• ChainCatcher report on the V4.1 Flash launch announcement, 10 September 2026