DeepSeek Just Made Its 75% V4-Pro Price Cut Permanent. Input Is $0.435 a Million Tokens. Here's When the Self-Host Math Finally Flips.
The 75% discount on DeepSeek's flagship V4-Pro was supposed to expire today, May 31, at 15:59 UTC. On May 22, the team said it isn't expiring. The cut is permanent. New API pricing: $0.435 per million input tokens, $0.87 per million output, and $0.003625 per million on cache hits. That last number isn't a typo. Cache reads cost about a third of a cent per million tokens.
Set aside the price-war drama for a second, because the interesting part isn't that DeepSeek undercut everyone. It's why they say they can hold the price. This isn't a loss-leader they'll claw back. It's an efficiency gain they're passing through.
Why "permanent" is the load-bearing word
A promotional discount is a bet on volume: sell below cost, hope you lock in users, raise prices later. Making it permanent is a different claim entirely. DeepSeek's framing is that V4-Pro was engineered for cheap long-context inference: roughly a quarter of the per-token compute and about a tenth of the memory footprint of its predecessor on long contexts. If that's true, the low price reflects a low cost to serve, not a subsidy.
That distinction matters for you because it changes how much you can trust the number. A promo price is something you build on at your own risk; it can vanish. A price grounded in an efficiency gain is something you can put in a unit-economics spreadsheet and expect to survive the quarter. I'd still keep a fallback provider configured (more on that below), but the planning horizon is different when the vendor is passing through a structural cost advantage instead of buying market share.
What this does to the self-host argument
The classic indie pitch goes: open weights are free, so rent a GPU and stop paying per token. The math has always been more annoying than that, and this price cut makes it worse for the self-host side.
Run the rough version. A mid-range GPU box capable of serving a frontier-class open model runs you somewhere in the $1–$2 per hour range on the cheap end if you rent it, more if you want reliability and headroom. Call it $1,000–$1,500 a month for something you'd actually trust in production with a little redundancy. At $0.435 per million input tokens, that monthly box budget buys you somewhere north of two billion input tokens through the API before you've broken even, and that's ignoring the output side, where the API is also cheap.
Two billion tokens a month is a lot for a one-person product. To beat the API on cost, you need steady, high-volume, latency-sensitive load that keeps your rented GPU busy most of the day. An idle GPU is the most expensive way to run inference there is. Most solo operators have spiky, low-average traffic: exactly the profile where renting per-token wins and self-hosting bleeds money on idle capacity.
So the honest answer for most readers: at these prices, the breakeven for self-hosting just moved out past where your volume realistically sits. Renting tokens is the right default unless you have a specific reason it isn't.
When self-hosting still wins
It's not never. Three cases still favor running your own:
You have predictable, sustained volume: a product doing real, constant traffic where the GPU stays hot. You have a hard data-residency or compliance requirement that rules out sending prompts to a third party. Or you have latency needs that an external API round-trip can't meet, and co-locating the model with your app actually matters.
If none of those describe you, you're optimizing a cost that the API already made small.
The honest counter-take
Cheapest-per-token is not the same as best decision, and I'd be doing you a disservice to pretend it is. DeepSeek is a Chinese-hosted frontier model. Routing your prompts and your users' data through it is a data-governance call before it's a price call. For a personal project, that's probably fine. For anything handling customer data, EU or otherwise, "where does this inference physically happen and under whose jurisdiction" is a question you answer before you look at the price sheet, not after.
There's also continuity risk. A provider aggressive enough to make a 75% cut permanent is competing hard, and hard competition is where vendors disappear, change terms, or get geo-blocked. The cheap-token era is real, but the move that survives it is staying portable: build against an abstraction so you can re-point to another provider in an afternoon, not a month.
So here's what I'd actually do. Use the cheap API. Don't self-host unless your volume or compliance profile demands it. But never hardcode a single provider into your product, because the thing that makes this pricing war great for builders (vendors fighting to be cheapest) is the same thing that makes any one of them a bad single point of failure. Rent the tokens, keep the exit ramp.
Author
Lukas
@lukcombinator