Falling Token Prices Were an Assumption, Not a Law

There is a sentence buried in a lot of AI business models, and almost nobody has written it down:

Inference gets cheaper from here.

It has been a reasonable thing to assume. Prices have fallen consistently, sometimes steeply, for two years. Plans got built on it. Margins got quoted on it. Pricing got set on it.

The vendor most responsible for making that assumption reasonable is now raising prices for the second time in under a month.

What changed

On 6 August 2026, DeepSeek announced it will raise the price of its API services, describing the increase as "significant."

That is the entire announcement, and the gaps are the story:

• No figure.
• No new pricing.
• No effective date.

For context on what's moving: DeepSeek-V4-Flash currently lists at $0.14 per million input tokens and $0.28 per million output tokens.

This is the second move, not the first

The framing in most coverage is that DeepSeek just reversed course. That undersells it, and the timeline is worth getting right.

• May 2026 — the company moved V4-Pro from a limited-time 25% discount to permanently lower pricing. Direction: down.
• Mid-July 2026 — DeepSeek introduced peak/off-peak pricing, doubling the rate during peak hours (09:00–12:00 and 14:00–18:00 Beijing time). Off-peak stayed at baseline. Direction: up, for anyone whose traffic lands in the working day.
• 6 August 2026 — a broad "significant" increase, unspecified.

So the reversal did not happen yesterday. It happened in July, and most people read it as a scheduling tweak rather than a price rise, because that is what demand-management pricing looks like from the outside.

That matters for how you weight this. One increase is an event. Two in four weeks is a direction — and the second one arriving before the first has even settled suggests the pressure is on cost, not on strategy.

Why it matters

DeepSeek is not just another provider raising prices. It is the vendor whose pricing forced everyone else's pricing down. The reversal comes from the origin of the trend.

That changes the shape of the variable, and the shape is what matters:

Inference cost is now a number that moves in two directions, set by a party with no obligation to warn you.

If your unit economics assume it only falls, you don't have a forecast. You have an undeclared dependency on somebody else's business decision.

And be clear about what that decision is. Cheap inference was never a property of the technology. It was a strategy — buy share, set the market price, absorb the cost. Strategies end when their job is done or their funding changes. There is no law of physics keeping tokens cheap.

The July change is the sharper lesson, because it shows the increase does not have to arrive as a headline. Peak/off-peak pricing doubled a lot of people's real bill without a single list price changing. If you were only watching the published per-million rate, you missed it.

The missing number is the actual problem

A bad number would be easier than this.

Tell me prices are going up 40% on 1 October and I can model it: run the numbers, see the margin, decide whether to re-price, migrate, or absorb it. It's unwelcome, and it's actionable.

"Significant," with no date, gives you nothing to compute with. You can't price the risk. You can only price that it exists. Which means the only rational response is to reduce your exposure to it — because you cannot calculate your way to a decision.

DeepSeek has also not said whether the increase applies equally across V4-Flash, V4-Pro, cache-hit pricing and reasoning models. Those have very different roles in most people's stacks, so "significant" could mean a rounding error or a re-platforming, depending on which line moves.

Proof and limits

The announcement is reported by Bloomberg, TechNode and The Next Web, all dated 6 August 2026. That is press coverage of a company statement, not a published price sheet — the pricing docs still show current rates as of writing.

Current V4-Flash pricing, the May discount change, and the July peak/off-peak structure are DeepSeek's own published figures.

What nobody knows: the size, the new prices, the effective date, and which models are affected. All unannounced, which means anyone telling you what this costs you is guessing.

This post is perishable. Its argument rests on the absence of a figure. The moment DeepSeek publishes one, the analysis changes — and it changes to a different piece, not this one with a number added.

What to actually do

None of this requires knowing the number.

1. Put a provider adapter between your code and the model API. If you call a vendor SDK directly from business logic, switching providers is a rewrite. If you call your own interface, it's a config change. This is a day of work and it is the entire difference between having options and having a negotiation.
2. Run a real cost model on your second-choice provider now. Not a spreadsheet of list prices. Actual token counts from your actual workload, priced against the alternative.
3. Check your effective rate, not the list rate. The July change is the warning here — divide last month's bill by last month's tokens and see whether that number already moved while the published price didn't.
4. Find out which of your features survive a 2x inference cost and which don't. Some will be fine. Some are only viable at today's price, and you should know which before the price moves rather than after.
5. Instrument spend per feature, not just total. A total tells you the bill went up. Per-feature tells you what to do about it.
6. Re-read your own pricing page with "inference could get more expensive" in mind. Unlimited plans are a bet on a cost curve.

Four questions, honestly answered

What is confirmed. DeepSeek announced on 6 August 2026 that API prices will rise significantly, per Bloomberg, TechNode and The Next Web. Current V4-Flash pricing is $0.14/M in, $0.28/M out. In May 2026 the company made a 25% V4-Pro discount permanent. In mid-July it doubled peak-hour pricing.

What is still unknown. The magnitude, the new prices, the effective date, which models are affected, and whether other providers follow. Nothing about the impact is modellable yet.

Why it is worth watching. It is a sustained reversal from the vendor who set the direction — two increases in a month, from the company whose whole position was being the cheap one. Whether others follow determines whether this is one company's margin problem or the end of a two-year trend.

How it fits the larger ecosystem. It changes the calculus on open weights immediately. Self-hosting has always been a trade of capex and operational burden against per-token cost; when per-token cost is falling, the trade is easy to refuse. When it moves both ways, running your own becomes a hedge rather than an indulgence — which is a live option this week in particular, given what else shipped.

The part that stays true

The cheapest provider is a business decision somebody else can reverse. It was never a feature of the technology.

Anyone who kept an adapter between their code and their provider can move. Anyone who didn't is negotiating.

Could you switch providers this week if you had to? If the model API is called straight from your code, that's not a migration. That's a rewrite — and you'd be starting it on their schedule.

---

Sources

• Bloomberg — DeepSeek plans significant price increase for its AI services
• TechNode — DeepSeek plans significant API price increases
• The Next Web — DeepSeek warns of a 'significant' price rise
• DeepSeek API pricing documentation