Two models and a calculator: what V4 actually costs
DeepSeek has put two versions of V4 on the market with prices that read like a typo. V4 Flash charges $0.14 per million input tokens and $0.28 per million output; V4 Pro rises to $0.435 and $0.87. The figures are listed on API pricing aggregators and on OpenRouter, where both models are already being served.
To make it concrete: processing the whole of Don Quixote through V4 Flash costs less than a vending-machine coffee. Frontier US models charge several dollars per million output tokens; DeepSeek plays in cents. They are not competing in the same sport — and that is exactly what Beijing wants you to notice.
A million tokens of context: what you can actually feed it
The other headline number is context: one million tokens, with outputs up to 384,000, on both models, with a switchable reasoning mode. In plain terms: a mid-sized code repository, a full season of scripts or your company's entire documentation fits in one conversation, no chunking.
A caveat: a big window is not infinite memory or infallibility — models still get lost in very long documents, and quality at full context is precisely what needs testing this week. But the combination of a huge window and rock-bottom pricing changes which projects are viable for a small studio or an independent developer. Things that required corporate budgets a year ago now fit on a freelancer's card.
On July 24 the old API goes dark, and there is no way back
Here is the date that makes this urgent: on July 24, the day after tomorrow, DeepSeek stops accepting the legacy model names on its API. Anyone with applications pointing at the previous generation's aliases has two days to migrate to the V4 names; after that, errors. It is the definitive end of the transition.
If you are a developer, the migration is trivial — the API is OpenAI-compatible, you change the identifier and move on — but it has to be done. And if you simply use tools built on top of DeepSeek, do not be surprised by a hiccup on Friday: someone, somewhere, never reads the notices.
China's AI week: V4 today, Kimi K3's open weights on the 27th
This is not an isolated release; it is a coordinated calendar offensive. With V4 deployed and the old API shutting down on the 24th, the spotlight moves to the 27th: the date Moonshot promised to release the open weights of Kimi K3, the model that already rattled Silicon Valley without anyone being able to download it.
The six-day calendar looks like this: V4 Flash and Pro live since today, the 22nd; the legacy API shutting down on Friday the 24th; and, if Moonshot delivers, Kimi K3 weights downloadable on Monday the 27th. Three hits in under a week, each aimed at a different audience: the developer watching prices, the company running production apps, and the open-source community that has spent a month waiting for the download.
Meanwhile, on the other side of the board, GPT-5.6 promises to do the whole job at prices dozens of times higher than DeepSeek's. The question I keep asking, and you should too if you pay for an API: how much of that gap is real quality, and how much is habit? I plan to test it on real cases this week. If Kimi's weights land on the 27th, Chinese AI will have closed its best week of the year in five days. If they do not, that is a story too: the promise will be a month old and still in the air.
The conversation starts here
Sign in with a supporter account to comment. Sign in
Nobody has commented yet. Want to go first?