DeepSeek V4 rolls out time-of-day pricing: during two daily windows its tokens —input and output— cost double and drop off-peak. Inference is starting to be billed like electricity, and budgeting for AI shifts from 'how many tokens' to 'at what hour.'
Picture opening your AI provider's console at ten in the morning and finding out that the same call that cost you a dollar last night now costs two. There is no fine print, no hidden penalty: it is, quite simply, the hour. DeepSeek just announced exactly that. With its V4 model, the Chinese lab introduces time-of-day pricing starting in mid-July: during two daily windows —from 9:00 to 12:00 and from 14:00 to 18:00, Beijing time— the price doubles, and outside those peaks it drops back down. It is the first time a major lab has charged you literally by the clock.
## Uber, but for thinking
The mechanism is identical to the surge pricing you already know from Uber or from time-of-use electricity rates: when demand goes up, the price goes up; when it drops, it drops. In concrete numbers, the V4-Pro version goes from 6 to 12 yuan per million output tokens at peak (roughly 1.70 dollars at peak); V4-Flash, from 2 to 4. And here is the nuance worth getting right: the surcharge does not fall only on what the model generates. According to The Next Web, during peak windows **the price of input also doubles** —the tokens you send it. Off-peak, nothing moves. DeepSeek is not hiding it as a trick: it presents it as a price signal so that whoever can will shift their load to the valley hours and free up GPUs when everyone is fighting over them.
## Why a lab dares to do this
Behind the idea is physics, not marketing. Serving a large model is a capacity problem: there is a finite number of GPUs and, during Asian office hours, they all saturate at once. Charging more at peak and less in the valley is the most honest way of saying 'I have no room right now; come back later and it will cost you half.' It is the same logic by which power costs a different amount at three in the afternoon than at three in the morning. And the twist is striking because DeepSeek was coming from the opposite direction: it had just made permanent a 75% discount on V4, after a year of undercutting the competition's prices. Going from giving away tokens to charging by the clock, in a matter of weeks, says a lot about how expensive compute capacity has become.
## The bill no longer depends only on how much, but on when
For anyone building software on top of these APIs, the change runs deeper than it looks. Until now, budgeting for AI meant multiplying tokens by a fixed price. With dynamic pricing, that math turns into a scheduling problem: a batch job that no one is in a hurry for —reprocessing documents, generating embeddings, overnight summaries— can cost half if you schedule it for the valley hours instead of firing it off at midday. And the other way around: an integration that ignores the clock can double its bill without anyone noticing, because the code works just as well; only the statement changes. Inference is starting to behave like a utility, and utilities are managed with time-of-day discipline, not with hope.
## What it means for anyone building with AI
The lesson is uncomfortable but liberating: if your product lives off calling models, cost is no longer a figure you check at month's end, it is a lever you operate in real time. That demands two things that until yesterday were optional: cost guardrails that understand context —not just 'how many tokens,' but 'at what price right now'— and the ability to shift non-urgent loads toward the cheap hours. At NeuralOS we work from that angle: the cost guardrails and the credit system exist so a pricey call does not slip past you by accident, and the automations engine with durable waits lets you defer batch jobs to the window you choose without losing the execution thread. We do not promise to negotiate each provider's clock for you —that would be smoke and mirrors—; what we give you is the dashboard so that 'when' stops being a surprise on the bill and becomes a decision you make.