DeepSeek adjusted its API prices for the V4-Flash and V4-Pro models starting today, August 16 — an increase that, for V4-Pro, reaches roughly 4 times the previous rate. The peak-hour structure we covered here in July stays the same; what changed is the base price sitting on top of it.
What changed
According to reporting from InfoWorld, Engadget, Fortune, and the South China Morning Post, the new V4-Flash pricing took effect at 16:00 UTC today: $0.22 per million input tokens and $0.66 per million output tokens during off-peak hours — nearly double what it cost before (roughly $0.28 per million output tokens). During peak hours, the price doubles again: $0.44 for input and $1.32 for output per million tokens. V4-Pro saw an even larger jump, going from roughly $0.87 to $3.96 per million tokens at peak — nearly 4 times more. DeepSeek attributed the change to the need to "allocate resources more reasonably," as demand for the models has been outpacing available capacity.
What hasn't changed is the time window itself: peak hours remain 01:00–04:00 and 06:00–10:00 UTC — which is 10 PM–1 AM and 3 AM–7 AM Brasília time, as we showed here in July. Brazilian companies running workloads during local business hours still fall, in practice, into the cheaper window.
Why it matters
The key point isn't whether the price went up or down — it's that it changed abruptly, unilaterally, and tied directly to the vendor's own capacity constraints. That's different from the gradual, well-announced adjustments mature vendors typically make. A company that calculated its cost-per-token using July's price table and hasn't revisited that number is now operating on a budget estimate that's off by more than 100% for some use cases.
The impact for Brazil
The structural advantage we pointed to here in July — Brazilian business hours falling inside DeepSeek's cheaper window — still holds in relative terms. But the absolute amount a Brazilian company pays per token rose significantly in a single update, without much advance notice. For anyone who built part of an AI project's ROI case around one vendor's per-token price, this is a reminder that number isn't a constant — it's a variable that can change overnight, especially for newer models still calibrating capacity to demand.
Entercast's take
We've said before, covering ByteDance's scale race, that a listed price isn't the metric that decides a vendor choice — it's just one input. DeepSeek's repricing reinforces that point from a different angle: no AI vendor, however cheap it looks today, should be treated as a long-term cost constant inside your financial model. The recommended practice is to review actual cost per task periodically — not just the advertised price per token — and keep at least one validated alternative on hand, so a unilateral repricing mid-budget-year doesn't leave you exposed.