DeepSeek Turns Its Famously Cheap Tokens Into Peak-Hour Pricing

For most of 2026, the simplest thing you could say about DeepSeek was that it was cheap. Absurdly, market-distortingly cheap — cheap enough that people started talking about a “death zone” where pricier models had no reason to exist. From 16 August, that story changes. DeepSeek is moving its flagship V4 models to peak and off-peak pricing, raising rates by as much as fourfold, and presenting the whole thing as a scheduling convenience.

Here is the short version, because the company’s own framing works hard to obscure it. At 16:00 UTC on 16 August, the flat price you pay today is replaced by two rates: a peak rate during the busiest seven hours, and an off-peak rate for the other seventeen, set at half of peak. That sounds like a discount is on offer. It isn’t. Every new rate — including the cheaper off-peak one — is higher than the single price you pay right now.

None of this makes DeepSeek expensive by the standards of its rivals. It stays well under what Anthropic or OpenAI charge. But this site measures from the point of view of the person paying the bill, and from that seat the change is unambiguous: the same work is about to cost more, the increase is dressed as a saving, and it arrives right before the company reportedly goes public. That is worth looking at plainly.

The discount that is actually a price rise

DeepSeek’s changelog explains the move like this: “To allocate resources more reasonably, we will adopt peak/off-peak pricing, with off-peak prices set at half of the peak-hour prices, encouraging users to schedule their tasks based on actual usage.” Read quickly, it lands as a win for the careful developer — shift your jobs to quiet hours and pay half.

The sleight of hand is in the word “half.” Off-peak is half of peak, a price that did not exist until this announcement created it. It is not half of what you pay today. When you line the new rates up against the current flat price — the only comparison that reflects your actual bill — the discount evaporates and a straightforward increase remains. This is the same move we flagged when Claude Sonnet 5’s price rise came twice: a rise you can defend on paper, packaged so the headline number sounds like a concession to the customer.

What the numbers actually do to your bill

Strip out the framing and look at the table DeepSeek published. Prices are per one million tokens.

• V4-Flash output moves from a flat $0.28 today to $0.66 off-peak and $1.32 at peak — up to 4.7 times the current price.
• V4-Pro output moves from $0.87 to $1.98 off-peak and $3.96 at peak — roughly 4.5 times, at the top.
• Input tokens rise as well. V4-Flash cache-miss input goes from $0.14 to $0.22 off-peak (up 57%) and $0.44 at peak; cache-hit input more than doubles even off-peak.

So the best case — you confine every request to the cheaper 17-hour window — is still a rise of well over 100% on output for both models. The worst case, running during peak, is a bill four to five times larger than the one you get today for the identical call. For a model family that people adopted specifically because the tokens were nearly free, that is not a tweak. It is a repricing of the entire value proposition.

And it compounds with a problem we’ve written about before. Token pricing is already hard to translate into a real invoice, because you rarely know your token counts in advance. Add a clock to it — where the same job costs one thing at 09:00 and double at 07:00 — and forecasting your spend goes from difficult to genuinely fiddly. As one developer on Hacker News noted, DeepSeek’s reasoning-heavy models lean hard on exactly the token types that are going up, so “the impact on those price increases will be felt much stronger” than the headline figure suggests.

Surge pricing, borrowed from your electricity meter

Now the steel-man, because there is a real one. Time-of-use pricing is not a scam; it is how electricity grids, cloud spot instances and ride-hailing all manage scarce capacity. If a service is congested at predictable hours, charging more then and less otherwise genuinely can push flexible load into the quiet windows and keep the thing usable for everyone. DeepSeek says that is the goal, and the congestion it points to is documented: in the weeks before the announcement, users logged repeated “server busy” errors clustered around the top of each hour, the fingerprint of demand outrunning capacity.

Concede all of that, and two things still stand. First, a demand-smoothing scheme could have been revenue-neutral — cheaper off-peak, dearer at peak, averaging out to roughly today’s price. DeepSeek chose instead to lift the floor, so that even the incentivised, inconvenient, middle-of-the-night option costs more than the status quo. That is a pricing decision, not a physics constraint. Second, congestion is a capacity story, and capacity is something a well-funded lab preparing to raise money is expected to build. Passing the shortage to customers as a surcharge is one answer; it is not the only one.

The clock is set to someone else’s working day

The advice buried in the announcement — just move your work off-peak — assumes your work is movable. For batch jobs, overnight pipelines and asynchronous agents, it often is, and those users will adapt fine. But peak is defined as 01:00–04:00 and 06:00–10:00 UTC, and that window is not neutral. It maps neatly onto the Chinese working day, and 06:00–10:00 UTC is 07:00–11:00 in London: a European developer’s entire morning sits inside the expensive lane.

If you run anything interactive — a chatbot your users talk to, an assistant answering in real time, a coding tool someone types into during office hours — you cannot simply defer that traffic to 3am to dodge the surcharge. Your peak is your users’ peak. “Schedule off-peak” is sound advice for a nightly cron job and empty advice for a live product, and a lot of the developers who built on DeepSeek’s cheap tokens were building live products.

The cheap era was the promotion, not the plan

Zoom out and the shape is familiar. A new entrant prices below cost to win developers, becomes the reference point everyone else is measured against, and then — once the workflows, the integrations and the habits are in place — starts moving the number up. We’ve watched flat subscriptions quietly turn into meters on the Western side of the industry; this is the same current running through the token economy, and it does not stop at any border.

Keep the comparison honest, though, because it cuts both ways. Even after Sunday, DeepSeek is dramatically cheaper than the frontier: at $3.96 per million output tokens, peak V4-Pro undercuts Moonshot’s Kimi K3 at roughly $15 and Anthropic’s Fable 5 at around $50 by a wide margin. So this is not a story about DeepSeek becoming expensive. It is a story about DeepSeek becoming ordinary — giving up the position that made it interesting, which was not “cheaper than average” but “cheap enough to change what you’d attempt.” A model priced so low it obviated pricier rivals is a different product from one that is merely competitive, and the developers who built ambitious, token-hungry things on the first assumption are the ones now re-checking their sums.

The IPO context sharpens it. DeepSeek is reportedly preparing to go public, and a company telling that story to investors has every reason to demonstrate a path to margin. Raising prices on a product people already depend on is the most direct way to do it. We should be careful here: we can’t read intent, and DeepSeek’s stated reason — resource allocation — is plausible on its own terms. But the effect is not in doubt. The pricing that made DeepSeek the industry’s floor was always going to be temporary, and the bill for the discount era is now being presented to the people who took it up.

What it means if you build on DeepSeek

The practical steps are dull and worth doing before Sunday. Re-run your cost model against the published table now, separating flexible traffic from interactive traffic, and work out how much of your load genuinely falls in the off-peak window versus how much only looks like it does. If your usage is deferrable, batching it off-peak is a real saving relative to peak — just not relative to last week. And build in the assumption that this number will move again: DeepSeek’s own page reserves the right to adjust prices, and it has now shown it will.

The larger lesson is the one this site keeps arriving at. A tool being cheap is not the same as a tool being cheap to you, permanently. The token that costs nothing today is a marketing decision, revocable at the vendor’s convenience and on the vendor’s schedule. DeepSeek is still, by the raw numbers, a bargain against its Western rivals. It is also, as of 16 August, a good deal more expensive than the one you signed up for — and a reminder to price your dependence on any single provider as if the promotional rate will end. Because it usually does.