openFDA is free. It is not unlimited.

The pitch for building your own FDA monitor is real: openFDA is public, keyless, and free, so a weekend of scripting against the device enforcement and 510(k) endpoints looks like it should replace a subscription outright. It does — for about three days. Then a run returns nothing but errors, on the exact week a recall actually needs reading, and the failure has nothing to do with the query. It is the ceiling FDA published on day one, and the script never checked for it.

The number FDA actually publishes

An anonymous, keyless caller to openFDA is capped at 1,000 requests a day and 240 a minute, per IP address. Register — free, an email and a form, no approval wait — and the daily ceiling rises to 120,000, same per-minute rate. FDA states both numbers plainly on its own authentication page. The gap between the two is not a paywall; it is closer to a courtesy the agency extends by default to nobody, because it does not yet know who is asking. Answering that one question — who is this — is worth 119,000 requests a day.

Where the free tier actually runs out

A thousand calls a day sounds generous until a script re-derives instead of remembering. Re-pulling ninety days of enforcement reports every morning, instead of asking only for a recent window, spends the daily budget on redundant history before a single new document is read. Querying several endpoints — enforcement, 510(k), classification — on every keystroke of a live search box multiplies the cost of one person’s curiosity by however many times they typed. None of this is abuse. It is ordinary engineering that happens to be metered by a number nobody checked in advance, and it fails exactly when the data matters most: a busy week is a high-query week.

A 429 response is not a bug in your script. It is the agency telling you, correctly, that you asked for more than you said you would.

The practice, not just the number

FDA Radar treats the same free endpoints as a shared public resource with a stated ceiling, not an unmetered feed: every fetch identifies itself by name in its User-Agent rather than arriving anonymous, each run asks for a bounded recent window rather than re-deriving the full history from scratch, and a database-level uniqueness constraint makes a second run over the same window a no-op instead of a duplicate fetch. Retries land only on a 429 or a 5xx, with exponential backoff and jitter so parallel adapters don’t retry in lockstep — never on a 4xx, since asking the identical malformed question again does not fix it.

For a team building this in-house

Register the free key before writing the first query — it costs an email address and raises the ceiling by two orders of magnitude before any code runs. Bound every query to a date window instead of pulling full history on every run, and let a uniqueness check on your own side absorb the overlap rather than fetching once and trusting memory. And read a 429 as an instruction to slow down, not an outage to route around with a retry loop that ignores it.

That is regulatory intelligence and method — how one free government API is actually metered, and the operating habit that keeps a self-built monitor reachable — not a claim about what any specific finding from it means for your device. Per the rules we publish under, the same discipline runs under FDA Radar’s own sources, whether or not a subscriber ever sees the calls being made.