Everyone Said AI Would Only Get Cheaper. On Sunday, DeepSeek Raises Prices Up to 1,100%.

At 16:00 UTC on Sunday, August 16, DeepSeek's API gets more expensive. V4 Pro output tokens go from $0.87 per million to $3.96 at peak hours, more than four times the current rate. V4 Flash output goes from $0.28 to $1.32. Across the V4 family the increases run from 50% to more than 1,100%, depending on the model, the token type, and the hour of the day the request happens to land.

This is the company whose global reputation rests on having proved that near-frontier performance could be delivered at a fraction of Western cost. DeepSeek is a large part of the reason a generation of founders assumed inference would keep getting cheaper. It is now the one raising prices.

The irony is the least useful thing in this story. The mechanism is worth more, and so is a structural change that most coverage has filed as a footnote: DeepSeek is introducing peak and off-peak billing, with off-peak set at half the peak rate. Time-of-day pricing is what capacity-constrained utilities do. It is not what a company sitting on spare GPUs does.

Two directions at once

The same week ran both ways. OpenAI has been cutting prices on selected models, and the Financial Times has reported Anthropic positioning Claude Opus 5 at roughly half the price of Fable 5. Those are competitive moves, and the competition they respond to is precisely the cheap Chinese alternative DeepSeek pioneered.

So the American labs are cutting under pressure from DeepSeek, and DeepSeek is raising. Both can be true at once because they face different constraints. A provider with ample capacity and a land-grab strategy cuts price to take share. A provider whose product is in more demand than it can serve raises price to ration it. The two behaviours look opposite and derive from the same variable: how much compute you have relative to how much of it people want.

The consensus model of AI economics that most product teams have internalised over the past two years holds that inference prices fall monotonically and permanently, because silicon gets cheaper and competition is brutal. That model is not wrong so much as incomplete. It describes a phase, not a law.

What a 1,100% increase actually is

The headline figure is doing a great deal of work. A range of 50% to more than 1,100% means the top of that range applies to one line item under one set of conditions, not to a typical bill. Anyone modelling their own exposure should work from the token types they actually consume rather than the largest number in the press coverage.

The figures confirmed across DeepSeek's announcement and multiple independent outlets:

• V4 Pro output: $3.96 per million tokens at peak, up from $0.87. Off-peak is half that, at $1.98.
• V4 Flash output: $1.32 per million at peak, up from $0.28. Off-peak $0.66.
• Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC. Everything outside those windows bills at half the peak rate.
• The new pricing goes live at 16:00 UTC on August 16, 2026.

Two things follow immediately. First, the effective increase for any given workload depends heavily on when it runs. A batch job that can be shifted outside the peak windows absorbs roughly half the pain. Second, the off-peak rate is not a discount in the ordinary sense. It is the peak rate divided by two, which makes peak the reference price and off-peak the concession, rather than the other way around.

It is also worth being straightforward about scale. Even after the increase, DeepSeek's models remain inexpensive relative to several frontier alternatives. A company is entitled to charge what its product is worth, and a price that was plausibly below cost during a land-grab was never a promise. This is not a betrayal story, and framing it as one obscures the part that matters.

Peak pricing is a capacity signal

DeepSeek's stated rationale is that peak and off-peak pricing lets it allocate resources more reasonably, and encourages users to schedule their tasks based on actual usage. That is a demand-management explanation, and demand management is what you do when supply is fixed in the short run. You do not need to ration something you have enough of.

That reversal is the most informative fact available here. A company that publicly forecast falling prices contingent on new hardware, and then raised prices instead, is telling you something about either the hardware timeline or the demand curve, and quite possibly both. Trade coverage has characterised the increase as demand straining capacity, which is consistent with the pricing structure the company chose.

Be careful about what this does not establish. It does not prove the Ascend 950 rollout has failed; demand growing faster than a perfectly successful rollout produces the same outcome. It does not tell you whether DeepSeek is profitable. And a company introducing time-of-day pricing is not a company in distress — utilities that price this way are usually well run. The signal is about scarcity, not about health.

The repricing risk nobody was modelling

For anyone whose product margin is a function of somebody else's token price, this week should change a line in the financial model. The assumption that input costs decline every quarter has been quietly load-bearing in a lot of business plans, and it has just taken a visible hit from the most credible possible source — not from an expensive incumbent defending margin, but from the cheap challenger everyone was benchmarking against.

The practical consequences are unglamorous and worth doing anyway.

• Model portability is a margin decision, not an engineering preference. If switching providers takes you a quarter, your supplier already knows that.
• Know your token mix. The increases were not uniform across input, output and cached tokens, so a blended assumption will misprice your exposure in one direction or the other.
• Workloads that tolerate latency should be scheduled, not merely sized. Off-peak batching is now worth real money with this provider, and plausibly with others before long.
• Optimising for cheapest-per-token selects for the provider most likely to reprice, because the lowest price is the one with the least margin defending it.

The broader point generalises well past DeepSeek. Cheap inference during a land-grab is a commercial strategy a provider can end unilaterally, on a few days' notice, without breaching anything. Peak and off-peak billing is the first widely visible case of an AI provider treating compute as a scarce, time-varying resource rather than an abundant commodity. It will not be the last, and the products built on the assumption of abundance are the ones that will find out first.