Thirty Cents a Million: DeepSeek V4.1 Flash and the Collapsing Cost of Intelligence

What This Changes for Someone Like Me

Okay, the personal part, because a price collapse only matters if it changes what a person can actually do.

In my post on whether indie developers can still make money, I argued the year was never the bottleneck — you were. I still believe that, and cheap compute doesn't fix it. Distribution is still king, the graveyard is still full, and a thin wrapper around a cheap model is still a thin wrapper. In fact, when intelligence gets this inexpensive, the gold rush gets worse: the wrappers multiply, because now every wrapper costs pennies to build too. If the only thing your product does is forward tokens, congratulations — your product is a rounding error on someone else's price list. The moat was never the model, and now it's even more obviously not the model. The moat is still the workflow, the data, the audience. That part of the ledger hasn't moved an inch.

What moved is the other side of the ledger. The side that used to be my excuse.

I run on a $0 stack — Cloudflare Pages, Supabase, free tiers all the way down — and for two years the honest caveat I attached to that stack was: "and then there's the AI bill, which does not have a free tier." That caveat just shrank until it stopped mattering — for my kind of usage, anyway. A million tokens of frontier-adjacent, vision-capable, tool-calling intelligence for thirty cents means the expensive ingredient in every indie project I've been putting off just became the cheapest one. Reading my whole codebase: fifteen cents off-peak. My profile — drafts, extractions, a proofreading agent puttering through a few million tokens a month — lands where the light users land: a couple of dollars a month. Being wrong: a fraction of a cent. And that last one is bigger than it sounds, because my whole playbook — lose cheap, learn fast — only works when losing is actually cheap. It just got its cheapest edition ever.

So here's what I'm actually doing with it. The project I'd been pricing out for months — a little tool that reads an entire project, understands the whole history of it, and answers questions like a senior engineer who never sleeps — I'd budgeted it like it was 2024. This week I rebuilt the cost model and it fit inside my coffee budget. One person, one weekend, one gigantically long context. Things that used to need a funding round now need a Saturday. "Small, sharp, finished" just got a superpower, and I intend to use it before the next keynote changes the numbers again.

Because here's the one thing I'm sure of, after this fortnight: the numbers will change again. That's the new rhythm. Flagships ship on Mondays, get retired by their own siblings on Fridays, and the price of thinking falls through the floor somewhere in between. If you build on any of it, build like the ground moves — because it does.

And underneath the numbers, the feeling I keep landing on is simpler than any spreadsheet: the excuse is gone. For two years the honest asterisk on every plan I made was the AI bill — the line item that didn't have a free tier, the one that made me close the tab before I started. This month somebody deleted the asterisk. When a million tokens of frontier-grade intelligence costs less than the coffee I drink while the download bar fills, the only things left between me and a shipped product are the idea and the work — and the work was always the part that mattered anyway. (´･ᴗ･`)

For the first time, the expensive part of building software is the idea. Everything else just got cheap, and I plan to spend accordingly.