Two Dollars a Day to Brief a Nation

I’m going to tell you what the Citizen’s Daily Brief costs to run, because the number is small enough to be surprising and because the surprising part isn’t the number. It’s what the number does and doesn’t tell you about how hard the thing was to build.

The Bill

LLM inference: about $60 a month. This is the only line that isn’t free, and nearly all of it is the daily pipeline: about $2 per run every morning, Monday through Saturday, plus a heavier Opus pass on Saturday night — the weekly assessment that carries Sunday. The rest of the stack rounds to nothing.

Compute: $0/month. The pipeline runs on GitHub Actions. A run takes about twenty minutes, most of it spent waiting on the model, and six runs a week plus the weekly pass comes to roughly six hundred minutes a month — well inside the free allowance. Nothing to pay.

Database: $0/month. Supabase’s free tier is more room than the CDB needs. Six tables, a few hundred rows of brief items a month, trivial query volume. It’ll want a paid tier eventually as the archive grows, but “eventually” is measured in years.

Web hosting: $0/month. A static site on Netlify’s free tier, rebuilt once a day. Serving pre-rendered HTML off a CDN is the cheapest thing in the world.

Total: about $2 a day, call it $60 a month. Weekday mornings are a flat couple of dollars off the metered bill; Saturday runs hotter when the weekly pass fires, and that one is still an estimate. The entire system — ingesting from thirty-four sources across the political spectrum, clustering the day’s stories, scoring their significance by how many distinct editorial perspectives cover them, synthesizing structured assessments with confidence and agreement signals, checking its own output, and publishing a briefing document — costs about what you’d pay for a cup of coffee.

Two Models, and Only One of Them Is Expensive

The reason the bill stays small isn’t a trick. It’s a decision about which model does which job.

The daily brief runs on Claude Sonnet, the fast, cheap tier. Everything it does is bounded: cluster today’s sources, score them against a fixed five-dimension rubric, synthesize each item into a schema I defined, check the result against the sources it came from. Constrained, repetitive, well-specified work, run dozens of times a morning, six mornings a week. Speed and price matter there, and the format itself is doing half the reasoning — when the shape of the answer is nailed down in advance, the model has less room to wander and less need for raw horsepower.

The weekly assessment runs on Claude Opus, which is slower and several times the price per token. That job is the opposite shape: read the whole week and say something genuinely synthetic about it, in long form, with no schema holding its hand. It’s the one task in the system that actually needs the bigger model’s judgment, and because it runs once a week, I can afford to give it.

There’s a cheaper tier below Sonnet, too. Claude Haiku runs about a third of Sonnet’s price, and if the only goal were the smallest possible bill, the daily brief would live on it. It doesn’t, on purpose. The daily analysis is where the product’s credibility sits — the significance calls, the assessments a reader is meant to trust before breakfast — and Haiku, capable as it is, lands just under the bar I need there. So I pay Sonnet’s premium for a floor I trust, and save Opus for the once-a-week job that earns it. The smallest bill was never the target; the smallest bill I’d stake the brief on was.

The same logic runs upward. There’s a tier above Opus, too — Claude Fable, built for long-horizon, many-step reasoning — and a fixed-shape daily brief is close to the last thing that needs one. The day I reach for it will be because the weekly assessment outgrew Opus, not because a bigger model makes a better product.

That’s the entire cost-control strategy, and it isn’t an AI strategy. It’s the oldest instinct in operations: you don’t run everything on the biggest box. You right-size the engine to the load — the cheap tier for the high-volume constrained work, the expensive one for the rare call that’s worth it. A daily run costs two dollars partly because most of it never touches Opus. Treating the models as infrastructure means provisioning them like infrastructure, per job, by weight.

What Actually Got Cheap

Five years ago I couldn’t have built this alone. Not because the architecture is hard — it’s a pipeline, a database, and a static site — but because the analytical core would have needed a room full of human analysts or a custom ML system that took months to train.

That specific capability collapsed in price. Taking thirty structured records, grouping them by topic, weighing their collective significance across defined dimensions, and emitting a structured assessment used to be a research project. Today it’s an API call that costs about two dollars a run. The raw reasoning got cheap.

Here’s the part I want to be careful about, because it’s easy to draw the wrong conclusion. Cheap to run is not the same as easy to build, and the invoice hides the difference completely.

Anyone can make these API calls. What took the months was everything the bill doesn’t show: learning what to ask, in what order, what counts as a valid answer, and — mostly — what to throw away before it reaches a reader. The pipeline that referred to President Trump as President Biden, that renamed the mayor of New York and changed his gender mid-paragraph, that swapped one pope for his predecessor inside a single sentence — every one of those was a lesson that turned into a check, a rule, a rewritten prompt. (I wrote that story up separately, in When the LLM Isn’t the Product.) None of it shows up as a line item. All of it is the reason the output is worth reading.

So the bill is low because inference got cheap. The product is defensible because the judgment stacked on top of that cheap inference didn’t get cheap at all. If the whole thing were really “three API calls and a few hundred lines of Python,” it wouldn’t be worth much — and it would be easy to clone. The API calls are the commodity. The accumulated judgment about how to use them is the product.

The PDB Comparison

The real President’s Daily Brief comes out of the Office of the Director of National Intelligence, drawing on the whole intelligence community. The budget isn’t public, but the community’s total runs around ninety billion dollars a year. The PDB is a sliver of that, and a sliver of ninety billion is still an enormous amount of money.

I’m not comparing the CDB to the PDB on quality — that would be absurd. The PDB draws on classified sources, satellite imagery, signals intelligence, and human networks no public system can touch. The depth of thousands of trained professionals is not reproducible by an API call.

But the format — a structured assessment with explicit confidence, shown reasoning, source attribution, delivered as one finite daily artifact — turns out to be the part that carries the value for a public reader. And the format is reproducible. The intelligence community never had a reason to notice, because for them the expensive, hard part was always the gathering. For the CDB, the hard part was believing the format alone, pointed at public sources, would produce something worth reading. It does.

Why Free Matters

The daily brief is free, and it stays free. Not a freemium tease with the good stuff walled off — what you read is the finished brief, not a sample of a paid one. Five to nine assessed items with full trust signals, evidence panels, and source attribution, every day, no account required.

That’s possible because the cost to run it is about sixty dollars a month. At that number the product doesn’t need to squeeze anyone. There’s no tier to buy and nothing sitting behind a wall. Readers who want to fund the work can, the way public radio works, and twelve supporters at five dollars a month cover the entire model bill.

And that changes the design. When your operating cost is five hundred dollars a month, you need hundreds of subscribers, or ads, or a venture round, and every one of those pushes you toward engagement optimization and attention farming. When it’s sixty, you need a few people who find it useful. The CDB can afford to be boring — no infinite scroll, no personalization, no engagement metrics, designed for read-and-leave instead of read-and-stay — because it doesn’t have to strip-mine your attention to survive.

The Stack

Worth listing, because almost every line is free:

The only paid line is the inference, and it’s the cheapest thing on the invoice relative to what it produces. Everything else is free or a rounding error. That isn’t a monument to clever engineering — the architecture is ordinary. It’s a monument to how good free-tier infrastructure has quietly become. The cleverness, such as it is, lives in the prompts and the checks, and those are plain text files.

What This Means

The CDB exists because structured reasoning turned into an API call, and everything around that API call turned free. A product that would have taken a team and a budget five years ago now takes one person and a coffee-money bill.

I think this pattern repeats: an expensive analytical process becomes a cheap call, and the cheap call unlocks a product format that used to be uneconomical. The CDB is an early instance because the briefing format was already well-defined and the inputs were public and free. The same move is sitting there in any domain where structured reasoning was the bottleneck.

But the cheap call is the beginning of the work, not the end of it. The question worth asking isn’t “what can AI do” — it can do plenty, most of it impressive demos and mediocre products. The question is “what valuable format was too expensive to produce until now, and what would it take to produce it well?” The first half is a search. The second half is where the years go.

For the CDB, the answer was a daily intelligence briefing for the public. Two dollars a day. One brief a day. Everyone gets the same one. That’s the product, and it’s only possible because the economics of analytical reasoning collapsed while nobody was looking — and only worth anything because collapsing the cost of reasoning did nothing to collapse the cost of judgment.

The Citizen’s Daily Brief is a free daily intelligence briefing from Stalefish Labs.