DeepSeek V4.1 Flash Pricing: Your Workday Is Off-Peak

DeepSeek put V4.1-Flash on its pricing page on September 10 with two columns, peak and off-peak. Convert those peak windows out of UTC into American clock time and something useful falls out: not one peak hour touches the US working day.

The short version: DeepSeek’s peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday. That is 6pm to 9pm and 11pm to 3am on the West Coast, 9pm to midnight and 2am to 6am on the East Coast. A business working 8am to 6pm anywhere in the continental US is billed off-peak for every hour of every working day, and does nothing to earn it. Off-peak V4.1-Flash is $0.15 per million input tokens and $0.60 per million output, verified on DeepSeek’s pricing page on September 11, 2026. The advice circulating since August, move your AI work to off-peak hours, is already true for you by default.

When are DeepSeek’s peak hours in US time?

Here is the conversion, for September, while the US is on daylight time.

The first window opens at 6pm Pacific, roughly when the last van gets back to the yard. The second runs through the middle of the night. When the clocks go back in November the windows shift an hour earlier locally, to 5pm Pacific and 8pm Eastern, so the conclusion holds through winter. We checked that case specifically, because daylight saving is exactly what breaks a rule of thumb three months after somebody writes it down.

Why the windows sit there is no mystery. In Beijing time, 01:00 to 04:00 UTC is 9am to noon and 06:00 to 10:00 UTC is 2pm to 6pm: a Chinese office day with the midday break cut out. China does not observe daylight saving, so the windows are fixed year round. DeepSeek is smoothing demand across its home market, and American customers sit in the trough.

What is the off-peak discount actually worth?

Less than almost anyone assumes. Take a three-van home services company that wants every inbound call turned into a job summary and a draft follow-up text. Say 45 calls a day across 22 working days, so 990 calls a month. A six-minute transcript runs to roughly 1,200 tokens, and with a 400-token instruction block attached that is 1,600 tokens in and around 250 back out.

That comes to 1.58 million input tokens and 247,500 output tokens a month. At off-peak rates the bill is 39 cents. At peak it is 77 cents. So the entire value of perfect off-peak scheduling here is about 38 cents a month, or $4.63 a year. You could rebuild your week around DeepSeek’s billing calendar and not clear the cost of a sandwich.

The general rule is simpler than it looks. Peak is exactly double off-peak on every line, so the most that perfect scheduling can ever save you is your own off-peak bill. Spending $6 a month? The discount is worth $6. For the timing decision to clear $100 a month you would need to be spending $100 off-peak, roughly 381 million tokens, or about 9,360 calls a day. At that volume you are not a small business.

Is there a bigger lever than timing?

Yes, and it sits in the next column of the same table, largely unremarked. DeepSeek bills a cached input token at $0.003 per million against $0.15 for an uncached one. That is a 50x difference. The off-peak discount is 2x. Caching is worth twenty-five times more than timing, and it applies to the repetitive part of the work: the standing instructions and reference material you send with every request.

Not every workload is cacheable, since it depends on whether your requests share a long common prefix. But if the conversation with your developer turns to trimming AI costs, the first question is whether caching is on and structured properly. The hour-of-day question can wait, probably forever.

How does this compare to what you pay now?

For context, the rates on Anthropic’s pricing page today: Haiku 4.5 at $1 and $5 per million input and output tokens, Sonnet 5 at $2 and $10, Opus 5 at $5 and $25.

Run the same 990 calls through each and the monthly bill is 39 cents on V4.1-Flash, $2.82 on Haiku 4.5, $5.64 on Sonnet 5. Worth noticing: DeepSeek’s peak rate, the one everybody is trying to dodge, is still more than three times cheaper than the least expensive Claude model. The window you are told to avoid sits well below the floor of the nearest alternative.

The weights are on Hugging Face under the MIT license, which permits commercial use without royalties, the same footing we looked at when an Abu Dhabi institute published its models along with the training data. The model card lists 552 billion backbone parameters with 8 to 16 billion active, a one million token context window, and image input.

So what should a small business take from this?

The useful conclusion is that at small-business volume, the model is not your cost line. Every option above runs under $6 a month for a thousand calls processed. You cannot optimise your way to a meaningful saving on a number that small, and the hours spent trying cost more than the saving.

What costs money is the software wrapped around the model, and the time it takes somebody to set it up and keep it honest. We flagged peak hours as a structure worth watching in August, when four providers moved their prices in four directions in two weeks. Having now converted the windows and run the numbers, the honest update is that for a US small business it is a non-event. That piece’s advice stands, keep your prompts and data portable, and one item comes off the worry list.

Which is the good news. When raw capability costs cents, the constraint stops being price and becomes whether you can describe your own process clearly enough to hand part of it over. That puts the leverage with the owner who knows the work, not whoever has the biggest budget. We ran a version of this arithmetic when DeepSeek made reading a receipt cost five hundredths of a cent, and the answer has not changed shape.

Frequently Asked Questions

Do DeepSeek’s peak hours affect me if I use ChatGPT or Claude instead?

No. Peak and off-peak billing is a DeepSeek structure, and it applies to its API rather than to consumer subscriptions. If you pay a flat monthly fee for a chat product, or use a tool somebody else built on a model, your price does not change by the hour. Treat it as a leading indicator of where metered AI pricing may go, not as something touching this month’s invoice.

What does DeepSeek V4.1-Flash cost per million tokens?

Off-peak it is $0.15 per million input tokens on a cache miss, $0.003 per million on a cache hit, and $0.60 per million output tokens. Peak rates are exactly double: $0.30, $0.006 and $1.20. Those figures come from DeepSeek’s own pricing page as verified on September 11, 2026, and the same page lists a one million token context window with a maximum output of 384K tokens, higher than several third-party trackers currently report.

Should I switch my business to DeepSeek because it is cheaper?

Price is a weak reason to switch when the gap is single dollars a month, which is what the arithmetic above shows at typical small-business volume. Better reasons to evaluate it: the one million token context window, the MIT license if you want the option of self-hosting, and whether it does your specific job well. Test it on real work of your own first, and treat cost as a tiebreaker.

If timing barely matters, is it still worth sorting my AI work by urgency?

Yes, but for operational reasons rather than billing ones. Knowing which jobs must happen while a customer waits and which can run overnight tells you where a slow, careful approach is fine and where responsiveness genuinely matters. That improves how you design the work regardless of what anybody charges, and it survives any future pricing change.

Here is what we are curious about. If the model itself costs under a dollar a month, what is actually stopping you from handing it a piece of your workload: the setup time, not knowing where to start, or the worry about what it gets wrong when nobody is watching? Those three have very different answers, and we would like to know which one is yours.