The cheapest API in the world stops being the cheapest tonight

Here's the deal: DeepSeek posted a notice on August 13. Alongside the full V4 lineup release, it's adjusting API pricing and introducing peak and off-peak rates. Effective time is 16:00 UTC on August 16, 2026 — which means if you're reading this during the US workday, you have hours, not days.

Start with the numbers. V4 Flash output tokens were a flat $0.28 per million. They become $0.66 off-peak and $1.32 at peak. Cache-miss input goes from $0.14 to $0.22 off-peak and $0.44 at peak. That's a 136–371% increase on output and 57–214% on input, depending on when you call.

But InfoWorld's headline says some prices went up more than 10x. I went through DeepSeek's own pricing table line by line to find which one, and there's exactly one: V4 Pro cache-hit input tokens. Those go from $0.003625 per million to $0.044 at peak. That's 12.14x, or +1,113%. Nothing else on the sheet clears 5x. So the "10x" headline rests on a single row — and that row happens to be DeepSeek's most famous cost weapon, which is what makes this story more than a routine price update.

The timeline is worth holding onto. On August 6, DeepSeek warned that prices would rise "in the near term" and that the increase would be "significant," with no figures and no date. On August 13 the actual table appeared. Today it goes live. Ten days, start to finish. And barely three months before that, on May 23, this same company announced it was making its 75% V4 Pro discount permanent. That permanence lasted 85 days.

What DeepSeek was, and why price was the whole identity

DeepSeek is based in Hangzhou. Liang Wenfeng, who built the hedge fund High-Flyer, founded it in 2023 and still holds the large majority of the equity. When R1 shocked the industry in early 2025, the shock wasn't purely about capability. It was that comparable capability was being sold for a small fraction of what American frontier labs charged.

Ever since, DeepSeek's business model has been essentially one thing: it sets the price floor. The model didn't have to beat the frontier. "Ninety percent of the capability at three percent of the price" was an overwhelming combination for bulk batch processing, data cleaning, agent loops, and chatbot backends. Every time a US lab revised its rate card, DeepSeek was the reference point sitting off to the side.

The V4 lineup currently splits two ways. V4 Pro is the generally available flagship; V4 Flash is the high-throughput, low-cost tier. Both support a 1M-token context window and up to 384K output tokens. This V4 Pro GA release also shipped stronger agent capability and adjustable thinking effort (low/high/max). So this wasn't a naked price increase — the product moved at the same time the price did.

Demand had already blown past capacity. On OpenRouter's weekly ranking for July 27 through August 2, V4 Flash 0731 came in first with 7.22 trillion tokens, and Chinese models swept the top four slots. On OpenCode, V4 Flash processed 8 trillion tokens on August 1 alone — 5 trillion of that on free usage, 3 trillion paid. Eight trillion tokens in a day means one model on one platform exceeded OpenRouter's entire platform-wide daily average of roughly 6.6 trillion.

And there are no chips to carry that traffic. Liang Wenfeng's leaked remarks put a number on it: training a frontier model required roughly 200,000 Huawei 950 chips, and Huawei could supply 16,000. Huawei's total expected output for the year sits around 750,000 units, split across every Chinese AI company. Liang expected the shortage to persist at least three years and called it basically unsolvable. US export controls have effectively removed Nvidia's high-end silicon from the table, leaving Huawei Ascend as the only real option — and Ascend is short. V4 was optimized for Ascend from its April preview onward. That was circumstance, not preference.

Reading the new rate card line by line

Straight from the official pricing page. All figures are USD per 1M tokens.

V4 Flash Old (flat) New off-peak New peak Peak multiple
Input (cache hit) $0.0028 $0.007 $0.014 5.0x
Input (cache miss) $0.14 $0.22 $0.44 3.1x
Output $0.28 $0.66 $1.32 4.7x
V4 Pro Old (flat) New off-peak New peak Peak multiple
Input (cache hit) $0.003625 $0.022 $0.044 12.1x
Input (cache miss) $0.435 $0.66 $1.32 3.0x
Output $0.87 $1.98 $3.96 4.6x

The widely reported "50% to 1,100%" range comes entirely out of these two tables. The floor is V4 Pro cache-miss input at off-peak ($0.435 → $0.66, +51.7%). The ceiling is V4 Pro cache-hit input at peak (+1,113%).

When peak actually falls matters enormously, and it isn't neutral. Per DeepSeek's docs, peak hours are 01:00–04:00 and 06:00–10:00 UTC — seven hours a day, seventeen hours off-peak. In China Standard Time that's 09:00–12:00 and 14:00–18:00, which is precisely the Chinese workday. If you're in Europe, peak lands mid-morning through early afternoon. If you're in Korea or Japan, it eats your entire working day. If you're on US Pacific time, almost everything you do falls in off-peak. The same rate card is a tax on some regions and a rounding error for others.

The genuinely painful change is what happened to caching. DeepSeek's context caching was the quiet weapon. Under the old V4 Pro rates, a cache hit cost 1/120th of a cache miss ($0.003625 vs $0.435). Under the new peak rates, it costs 1/30th ($0.044 vs $1.32). The caching discount got cut to a quarter of what it was. V4 Flash went from 1/50th to about 1/31st. InfoWorld reports DeepSeek maintains roughly 98% cache-hit rates against an industry norm nearer 90% — which means the workloads that leaned hardest on that advantage are exactly the ones absorbing the biggest increase.

Let me show it with numbers. Take an agent loop that replays a long system prompt over and over: 10M input tokens at 98% cache hit, 200K output, on V4 Pro.

  • Old: (9.8 × $0.003625) + (0.2 × $0.435) + (0.2 × $0.87) = $0.297
  • New off-peak: (9.8 × $0.022) + (0.2 × $0.66) + (0.2 × $1.98) = $0.744 (2.5x)
  • New peak: (9.8 × $0.044) + (0.2 × $1.32) + (0.2 × $3.96) = $1.487 (5.0x)

The better your cache discipline, the worse your increase. "Stuff the prompt and let caching absorb it" was the optimal strategy on DeepSeek, and it just became the most expensive one.

So is it still cheap? Let's do a flat comparison — one run of 1M input plus 1M output, using rates pulled from each vendor's own documentation.

Model Input / Output (per 1M) Cost of 1M + 1M
GLM-4.7-FlashX $0.07 / $0.40 $0.47
Gemini 2.5 Flash-Lite $0.10 / $0.40 $0.50
DeepSeek V4 Flash (old) $0.14 / $0.28 $0.42
DeepSeek V4 Flash (off-peak) $0.22 / $0.66 $0.88
DeepSeek V4 Flash (peak) $0.44 / $1.32 $1.76
GPT-5.6 Luna $0.20 / $1.20 $1.40
GLM-4.7 $0.60 / $2.20 $2.80
DeepSeek V4 Pro (off-peak) $0.66 / $1.98 $2.64
DeepSeek V4 Pro (peak) $1.32 / $3.96 $5.28
Grok 4.6 (under 200k) $2.00 / $6.00 $8.00
GPT-5.6 Terra $2.00 / $12.00 $14.00
Kimi K3 $3.00 / $15.00 $18.00

The answer splits. At the budget tier, DeepSeek lost the crown. V4 Flash costs $0.88 even off-peak, which is above Google's 2.5 Flash-Lite ($0.50) and Z.ai's GLM-4.7-FlashX ($0.47). At peak it hits $1.76, which is more expensive than OpenAI's own cheapest tier, GPT-5.6 Luna, at $1.40. For seven hours a day, the Chinese budget model costs more than the American budget model. That's the reality behind Sanchit Vir Gogia of Greyhound Research telling InfoWorld that "on paper, at peak, against the right comparator, DeepSeek's price advantage does disappear."

At the frontier tier it's still cheap. V4 Pro at peak runs $5.28, which is 38% of GPT-5.6 Terra ($14.00), 66% of Grok 4.6 ($8.00), and 29% of Kimi K3 ($18.00). Off-peak at $2.64, it's a little under a fifth of Terra. If you want near-frontier capability on a tight budget, DeepSeek remains a real option. What changed is the order of magnitude: "a fraction of a fraction" became just "a fraction." Alibaba's Qwen3.8-Max is widely quoted at $2/$6, but I couldn't confirm that from Alibaba Cloud's own documentation, so I left it out of the table.

Who gains from the increase

DeepSeek gains two things. First, load shaping. Off-peak at half the peak price is a design for pushing traffic around before it's a design for collecting more revenue. Seventeen of twenty-four hours are off-peak, so anything that isn't time-sensitive drifts there. The company's own phrasing in the notice is "more flexible workload scheduling." Second, unit economics. Processing 8 trillion tokens a day at $0.28 per million doesn't close no matter how well you optimize GPUs. Selling more was making the hole deeper.

DeepSeek's balance sheet is part of this too. Fortune reports the company is running a large fundraise while preparing for a possible listing this year. It is hard to write a prospectus around a business that loses money on every token it sells. This price increase is a product decision and a financial cleanup at the same time.

Western frontier labs are quietly delighted. For eighteen months, every pricing decision at OpenAI, Anthropic, and Google carried an unspoken "but DeepSeek sells it for this" in the background. When the reference point moves up 3–5x, defending your own rate card gets much easier. OpenAI can now truthfully say Luna undercuts DeepSeek Flash at peak on both input and output.

Chinese rivals get an opening. Z.ai's GLM family, Alibaba's Qwen, and Moonshot's Kimi all fish in the same water, and DeepSeek's price was simultaneously their ceiling and their floor. With GLM-4.7-FlashX at $0.07/$0.40, the budget segment already has a landing spot picked out.

The losers are the apps that built a business model on DeepSeek's price. Anyone selling "unlimited AI features for $9 a month" is in this bucket. When token cost moves from a flat $0.28 to somewhere between $0.66 and $1.32 depending on the clock, your cost of revenue doesn't just double or quintuple — it becomes unpredictable. A cost base that varies 2x with the time of day is about the worst possible match for a flat-rate subscription.

We've seen loss-leader pricing before — the endings differed

Selling below cost to take a market and raising prices later isn't new. Cloud, ride-hailing, and food delivery all walked that road. The outcomes split, and the dividing line is clear.

The ones that succeeded had a lock-in mechanism ready before they raised. Cloud providers are the model case. They sold compute cheap early, but databases, storage, identity, and networking stacked on top until leaving meant rearchitecting rather than switching a config value. Ask whether DeepSeek has anything like that and the honest answer is: not much. The API is OpenAI-compatible, so moving to another provider is a few lines of code. Compatibility was the weapon that drove adoption; now it's the exit ramp.

The ones that failed had price as their only differentiator. Users acquired on price are loyal to price. Raise it and they walk to whoever is next-cheapest, and on the way out they attach an "unreliable" label to your brand. DeepSeek's specific exposure here is that May "permanent" announcement. Raising a price you explicitly called permanent, three months later, means every future DeepSeek pricing statement carries a discount factor. Even if the company cuts prices next quarter, nobody is going to build a twelve-month plan on it.

That said, this case differs from the classics in one important way. Uber and WeWork burned capital, but their supply was effectively unlimited. DeepSeek's constraint isn't capital — it's physical silicon. Burning more money doesn't produce throughput if Huawei can't ship more Ascend parts. In that situation, a price increase isn't greed; it's rationing. If Liang's "at least three years" estimate holds, this isn't a one-off adjustment. It's the start of a new equilibrium.

And there's the Anthropic precedent. InfoWorld connects this decision to the same pressure behind Anthropic's April pricing move. When demand outruns capacity, suppliers eventually queue customers by price. The whole industry is walking away from the assumption that inference can be sold below cost indefinitely, and DeepSeek was the company that carried that assumption longest.

How the rivals counter

Z.ai (GLM) is the most direct beneficiary. GLM-4.7-FlashX is $0.07/$0.40 and GLM-4.7 is $0.60/$2.20 — clearly under DeepSeek V4 Flash off-peak ($0.22/$0.66) in the budget band. On top of that, GLM-4.7-Flash and GLM-4.5-Flash are free. Expect migration guides and possibly sharper pricing aimed straight at churning DeepSeek volume.

Google already wins in a segment it didn't have to fight for. Gemini 2.5 Flash-Lite is $0.10/$0.40, and batch processing halves that to $0.05/$0.20. DeepSeek can advertise its off-peak discount all it likes; it doesn't reach Google's batch tier. For any developer outside China shopping for a cheap API, Google is the path of least resistance right now.

OpenAI just acquired a marketing angle. GPT-5.6 Luna is $0.20/$1.20 against V4 Flash peak at $0.44/$1.32 — cheaper on both sides of the meter. A company that spent eighteen months absorbing "the Chinese models are far cheaper" now has a seven-hour-a-day rebuttal.

Moonshot (Kimi) and Alibaba (Qwen) compete on a different plane. Kimi K3 at $3/$15 isn't playing the budget game at all; it sells agentic coding capability. When V4 Pro climbs to $1.32/$3.96 at peak, the absolute gap remains large but the psychological one narrows. More teams will conclude that if nothing is truly cheap, they may as well buy the best.

Open-weight self-hosting may be the real counter. DeepSeek's own weights are published, so you can run the model yourself. The reason nobody bothered was simple: the API was so cheap that renting GPUs lost money on arithmetic alone. At $0.28 per million output tokens you didn't need a spreadsheet. At $1.32 the break-even point drops sharply. If your traffic is steady and predictable, it is genuinely worth rerunning that calculation this week. The hidden costs — GPU procurement, ops headcount, latency management — haven't gone anywhere, but the gap they have to overcome just shrank by a factor of four.

So what actually changes

If you call the DeepSeek API, you have two things to check before the cutover. First, pull your logs and see how much of your volume lands inside 01:00–04:00 and 06:00–10:00 UTC. Second, check your cache-hit rate. The higher it is, the worse your effective increase — a workload above 90% hit rate can see a real-world multiple well beyond the headline percentages.

If you run an AI product, your pricing page needs a look. Selling flat subscriptions while your input cost swings 2x by time of day makes margin unmanageable. Three practical responses: queue non-urgent work into off-peak windows, tier your routing so cheap requests go to a cheaper provider, or move your own plan to usage-based billing. The third hurts most and is the most honest.

If you're in enterprise IT, you now have a concrete argument for provider diversification. A price the vendor called permanent moved in three months. That's not a hypothetical risk anymore, it's a documented event. Putting an OpenAI-compatible routing layer in front of your inference calls turns the next pricing event from a code change into a config change.

If you invest in the Chinese AI value chain, there are two signals here pulling opposite directions. The bearish one is that the growth bottleneck is confirmed to be silicon rather than demand, and it showed up in the price sheet. The bullish one is that DeepSeek is finally aligning revenue with cost, which matters more than usual for a company preparing to list. What nobody knows yet is how much volume actually walks. Watch OpenRouter share through late August and September — drawing conclusions before that data lands is premature.

If you're weighing open-weight self-hosting, the premise of your calculation changed. The API used to win outright; now it depends on your traffic shape. Flat, all-day load with loose latency requirements tilts toward your own infrastructure. Spiky, bursty load still favors the API — which is precisely the spikiness DeepSeek is trying to flatten by selling off-peak at half price.

If you're a regular user, nothing visible changes today. The DeepSeek app and web chat aren't part of this adjustment. What may show up over the next few weeks is third-party apps that quietly ran on DeepSeek's API trimming their free tiers or reworking their plans. If an app you use suddenly gets stingier, this is a plausible reason why.

🥄 Three Things You're Probably Wondering

— So is DeepSeek just not cheap anymore? Half right. V4 Pro still runs at about 38% of GPT-5.6 Terra even at peak, so the frontier tier remains a bargain. But V4 Flash is no longer the price floor — Gemini 2.5 Flash-Lite and GLM-4.7-FlashX both undercut it, and at peak it costs more than GPT-5.6 Luna. The sentence "DeepSeek is always the cheapest option" expired today.

— Can't I just move everything to off-peak? In theory, and it depends brutally on where you sit. Peak is 01:00–04:00 and 06:00–10:00 UTC, which maps neatly onto the Chinese workday and lands squarely on East Asian and European working hours. Interactive work that needs a human watching can't be shifted; overnight batch jobs can. Seventeen hours of off-peak is real headroom, but which seventeen hours you get is an accident of your time zone.

— Is this a temporary adjustment or the start of a trend? Too early to say that flatly, but the evidence leans against temporary. Liang put the Huawei Ascend shortage at three years minimum, and export controls have closed the Nvidia alternative. If capacity can't grow, price has to suppress demand. There are counter-scenarios: Huawei output ramping faster than expected, DeepSeek's own silicon efforts landing, or churn coming in bad enough to force a reversal. Either way, after the May "permanent" cut, taking this company's pricing statements at face value is not the move.

References

Numbers are as of announcement and may change.