Builds That Cost Under a Dollar

Most "cheap AI projects" lists are really lists of free tiers. This one is a ladder of real builds made with Jev, the decision model from TypeSafe AI, where the builder published what the whole run cost, and it came in under a dollar. Every figure is as reported by the author, linked to the entry; none are official prices.

The full cost table, and the napkin method for budgeting your own project, live on what Jev builds actually cost. This page is the "what could I build for pocket change" tour, climbing one rung at a time.

One caution before the ladder: some receipts are whole-run totals and some are per-unit rates. We say which is which. Mixing them up is how a per-search price turns into a surprise invoice.

Rung one: fractions of a tenth of a cent

• Music intelligence audit: 3,274 songs across 114 genres, about $0.0001 total, as reported. The cheapest whole job in the catalog.
• AI slop detector: 35 tells checked on a page in 243 ms for $0.00015.
• Post scoring with SuperX: 61 questions about one draft for $0.0004.
• 400 companies matched to one candidate: 12 seconds, $0.0005.
• Second-hand shopping agent: a 406 ms decision per listing, $0.00085 reported for the run.

Per-unit, not totals: computer use on macOS at about $0.0002 per step, and a context-aware meme picker at $0.0003 per search.

Rung two: under a cent

• Every's editorial vibe check: 1,709 judgments over 37 documents for under a cent.
• Flight search with Browser Use: a full browser-agent run for about $0.004.
• A 26-sheet plan set in 2.9 seconds: $0.0052.
• Doomscroll Filter: three days of a niche's posts, eight questions each, $0.007.
• Subway Surfers, 50 games at once: under a cent, per the builder.

Per-unit: the YouTube sponsor skipper at about $0.005 per video.

Rung three: under a dime

• Which exam questions come back: 377 questions in 80 seconds, $0.0256.
• 500 emails for 3.5 cents: the receipt that launched a thousand inbox experiments.
• Fraud detection with Jev and Kimi K3: 96 of 100 correct for about $0.07, including the frontier model handling the unsure cases.
• 1kpapers: 1,018 research papers classified for $0.08.
• 700 leads scored in 40 seconds: $0.09.

Rung four: under a dollar

• 3,282 posts, eight questions each: $0.1282 for an entire X archive.
• 723 ads, 30 personas: 21,690 decisions for 22 cents.
• Every YC company interviews his resume: 6,245 companies in 25 seconds, $0.37 total.
• Tocsin: 22.8 million log lines for $0.64, against an estimated $1,120 for running an LLM over every line, per the author.
• 100,000 viral posts in 20.4 seconds: $0.67, the top rung.

Where the dollar runs out

For honesty's sake, the entries that don't fit here. Continuous play is where costs add up: Jev plays Doom is reported at about $7 an hour. And a serious adversarial evaluation, jev-evaluation, ran 123,805 requests for $12.69. Neither is expensive, but "pennies per job" becomes "dollars per hour" once a loop never stops.

For what a single call runs, see how much Jev costs per request. For keeping a cheap pipeline cheap, how builders measure cost covers the field methods.

Frequently asked questions

What is the cheapest Jev build in the directory?

The music intelligence audit, at about $0.0001 for 3,274 songs as reported. Several single-check tools, like the slop detector, come in at a fraction of a tenth of a cent per check.

Can I build a useful AI project for under a dollar?

The builds above suggest yes for decision-shaped work: classification, scoring, triage and ranking. Anything that needs generated text needs a chat model for that part, which changes the math.

Are these costs guaranteed?

No. They are builder-reported, at launch-period pricing that can change; docs.typesafe.ai is the source for current rates. Use the pricing page's napkin method to estimate your own.

Why do costs vary so much between builds?

Volume, questions per item and input length all multiply together, and some builds report totals while others report per-unit rates. Compare like with like before drawing conclusions.