Cache-hit input tokens on DeepSeek-V4-Pro just went from $0.003625 per million to $0.044 at peak. That's, by this article's reading, the 1,100% headline — a number that reads like the price war's obituary, cheap Chinese AI finally breaking ranks.
Read cold, it's the reversal of the story we ran two weeks ago. Last issue: OpenAI cut Luna prices by 80% off-peak and the deflation narrative looked unstoppable. This week: DeepSeek, the poster child of undercutting everyone, raises rates by up to eleven times.
But here's the number nobody quotes. Peak hours on DeepSeek's own clock are 01:00–04:00 and 06:00–10:00 UTC. That leaves 17 of every 24 hours at half price — off-peak.
And Western buyers mostly pay those off-peak rates. So the shock number describes a window most developers outside China won't be awake to touch.
Look at what actually repriced. V4-Pro output tokens rise to $3.96 per million at peak and $1.98 off-peak, up from a flat $0.87. That off-peak figure is real, but not eleven-fold.
For context, Anthropic's Fable 5 charges $50 per million output tokens. "Cheap AI" did not become expensive AI; it became slightly-less-absurdly-cheap AI, on a schedule.
The 1,100% only appears in one specific cell: cache-hit input tokens, the tiny fraction where apps reuse stored prompts. That's the mechanism worth understanding, because it's where DeepSeek's real edge lived.
Sanchit Vir Gogia at Greyhound Research puts it precisely: DeepSeek's roughly 98% cache-hit discount, against an industry norm nearer 90%, "has kept its measured cost per task at about 60% below Luna," even after OpenAI's Luna cut. "The schedule re-prices exactly that mechanism," he says. "The cache is where the advantage genuinely erodes."
Watch how the edge narrows on Flash. Gogia's math: Flash's advantage over Luna drops from roughly sevenfold to threefold off-peak, and to 1.4x at peak. The cache repricing doesn't kill the lead — it thins it, and only during the hours few Western developers use.
Which is why the honest read is not "prices are going up." It's that DeepSeek built a pricing machine to move load off its most congested hours. The company said as much: the tiered structure exists "to allocate resources more reasonably," to push developers toward "less congested periods."
DeepSeek built a pricing machine to move load off its most congested hours — supply rationing dressed as repricing.
Gogia states the tell plainly. The new schedule "hits DeepSeek's home market hardest and its export market lightest." Chinese daytime usage reportedly lands squarely in the peak windows; Western buyers largely sit in the cheap ones.
That reads less like repricing than, in this article's interpretation, supply rationing dressed as repricing — a way to squeeze the home market for compute while keeping the export market's rates soft.
The motive lines up. KB Securities argues the primary aim "is not to increase ARR" but to "secure computing power" for a reported 1GW data center in Inner Mongolia. This is a fundraising and capacity move, not a margin move.
And KB draws the ceiling. To match cheap U.S. models, V4-Pro would need to rise more than threefold; the actual increase, they expect, will be "no more than double." The 1,100% is a corner case, not the trend.
Why can't DeepSeek hike further? Because the model is open-weight. Researcher Ryu Ho-seung's warning: if DeepSeek raises API prices too far, "other AI service providers could use DeepSeek models to offer tokens at a lower price."
That's the leash. DeepSeek's own weights are the arbitrage. Push the sticker price too high and third parties undercut you with your own model — the price ceiling is enforced by your own product.
So the deflation story survives one provider's correction. KB reads the whole move as a shift "from an ultra-low-price strategy to a low-price strategy," not the end of price competition — and expects Chinese oversupply to persist.
Benedict Evans pushes the same direction structurally. He argues token prices are unstable and headed toward low-margin commodity infrastructure, the way fiber overbuilt in the Dotcom bubble — the current spike is a supply crunch, not a floor.
Mark Tauschek at Info-Tech frames the hike as plain physics: "simple supply and demand: when demand goes up, pricing goes up, because supply becomes constrained." He notes Anthropic raised prices in April for the same reason. Rising demand meeting scarce compute is not the death of cheap tokens; it's the tide going out and coming back.
Now the operator takeaway, because this is where it matters if you build on these APIs.
Your unit economics no longer depend on a sticker price. They depend on two variables you now have to actively manage: what fraction of your clock runs off-peak, and what fraction of your input tokens hit cache.
Schedule arbitrage is a real lever now. Work that can wait — batch jobs, evals, overnight agent runs — moves into the 17 cheap hours and pays half. Gogia's line: "timing becomes an economic variable, and work that can wait moves into the cheap hours."
Cache management is the other lever, and it's the fragile one. The 98% cache-hit discount is what put DeepSeek 60% below Luna per task; the new schedule taxes exactly that. If your architecture reuses prompts heavily, your bill just got more sensitive to when those hits land.
The two variables interact. A cache-heavy workload running in peak hours is now the worst case; a cache-light batch job off-peak barely moved. Same model, same task, wildly different cost depending on the clock and the hit rate.
Tauschek's warning about model routing lands here. Organizations that ran usage flat and paid per seat are hitting "sticker shock akin to that of the early cloud days." The teams that win are the ones treating tokens like cloud compute — routed, scheduled, cached deliberately.
CIOs should read the schedule, in Gogia's phrase, with "relief and unease." Relief because the bill is largely schedulable. Unease because a supplier that has learned to reprice this granularly can reprice again.
That's the durable lesson. The subsidy era isn't ending — it's getting a clock and a cache-hit rate bolted onto it.
Stop reading the headline number. Start reading the schedule, because that's now where your margin lives.