DeepSeek is raising API prices. The cheap-cloud-AI bet just moved.

For two years the counterargument to running your own model was a single number: the API is so cheap it is not worth the trouble. DeepSeek made that argument better than anyone, undercutting every Western lab and moving a staggering volume of tokens at rates that looked too low to be real. This week DeepSeek confirmed they were too low to be real, and the DeepSeek API price hike it just signaled is the clearest sign yet that the cheapest-API pitch was borrowing against a subsidy all along.

What did DeepSeek actually say?

That prices are going up significantly, that developers should plan accordingly, and that the specific new rates will follow in an official notice. DeepSeek posted the warning on its own pricing page, with no figure and no effective date: “We plan to raise the overall pricing for DeepSeek API services in the near future, with a significant increase expected.” It was reported by Bloomberg and picked up widely, and it lands just after DeepSeek introduced peak-hour surge pricing, so this is the second time in short order that the cost of using it has moved.

For reference, DeepSeek-V4-Flash is priced at $0.14 per million input tokens and $0.28 per million output on DeepSeek's own rate card as of this writing. That is the rate a lot of tools and side projects quietly standardized on, precisely because it was so far below everything else.

Why is the cheapest model on the market raising prices?

Because the cheap price created demand the cheap price could not fund. The reporting around the announcement describes V4-Flash usage growing explosive enough to overwhelm DeepSeek's available compute, which is what happens when you price a capable model below cost and the internet notices. Rock-bottom rates do not pay for the hardware that serves them. This is not a story about a greedy vendor. It is a story about arithmetic catching up with a subsidy.

A price set below cost is a marketing budget, not a market rate. It lasts exactly as long as the vendor decides to keep spending it.

Is DeepSeek still cheap? Yes. That is not the point.

Even a tenfold increase would likely leave DeepSeek cheaper than most closed frontier models. Push the $0.14 input rate up ten times to $1.40 per million and it still lands far under the frontier tiers: Anthropic's Fable 5, for one, runs $10 per million input and $50 per million output. If your only question is whether this is the cheapest API, DeepSeek probably stays the answer.

But that framing hides the actual risk. The number you wrote into your cost model was chosen by DeepSeek, and in one month it moved twice, with no figure attached to the second move. A provider that can reprice on short notice, for reasons entirely outside your control, is a provider whose costs are theirs to set and yours to absorb. Cheap-but-someone-else's-call is a different thing from cheap-and-yours.

This is the same trap as metered tools, one layer down

We have written before about the cost wall that metered coding tools hit, when even Uber capped its own engineers after burning a year's AI budget in four months. That was the tooling layer: Copilot, Cursor, and the rest turning flat subscriptions into meters. This is the layer underneath it, the raw model API that those tools and your own scripts call. When the floor moves, everything built on it moves too.

The pattern is identical at both layers. A price is set low to win the market, developers architect around it, and then the price finds its real level once the usage is locked in. The only durable defense is to not be standing on a floor that someone else owns.

The out the metered tools never had: you can download this one

DeepSeek-V4-Flash is MIT-licensed, and the weights are on Hugging Face. That is the difference that matters here. When it shipped on July 31, the local runtimes caught up within days: llama.cpp added the architecture, ExLlamaV3 and vLLM followed, and community GGUF quants appeared shortly after. A model you can pull is a model whose price cannot be raised on you, because there is no meter between you and your own hardware.

The honest caveat is hardware. V4-Flash is a large Mixture-of-Experts model with a 1M-token context, so running the full checkpoint is a server-class job, not a laptop one; reported 3-bit builds want somewhere around 110GB of RAM. “You can own it” and “you can run it on the machine in front of you” are still two different claims. For most developers the practical shape is a mix: a smaller local model, a Qwen or Llama coder in the 7 to 32B range, handling the routine work at zero marginal cost, and a cloud API reserved for the occasional hard problem where a frontier model earns its rate.

What to actually do about it

Stop treating a subsidized price as a foundation, and put the parts you control back under your feet. Four concrete moves:

• Use bring-your-own-key, not a reseller. If you reach DeepSeek through a tool, make sure you are paying DeepSeek directly at its published rate, with no platform markup stacked on top of a price that is already climbing. BYOLLM means the only bill is the provider's.
• Keep a local model loaded for the routine 80%. Most coding requests are not frontier tasks. A local model via Ollama or LM Studio answers them at zero inference cost and never sends a repricing email.
• Self-host the weights if a specific open model is core to you. Where the license allows it, and MIT does, running the weights yourself turns a variable API bill into a fixed hardware cost you control.
• Do not architect a business on a floor you do not own. Build so you can switch models and providers without a rewrite, so the next repricing is an afternoon of work, not a crisis.

DeepSeek did nothing wrong here. It priced to win, won, and is now pricing to survive the winning, which is what any company would do. The lesson is not about DeepSeek. It is that a price someone else subsidizes is a favor, not a fact, and favors get withdrawn. The parts of your stack that run on your own hardware, with your own keys, are the parts nobody can reprice while you sleep. That is the entire case for local-first, delivered this week by the cheapest cloud model on the market.