Developer reference · live data

What timeout should you set for each free API?

Averages hide the slow requests that break your app. These are the median, 95th and 99th percentile response times of 25 free public APIs over the last 7 days, measured every 10 minutes from a server in Sweden, plus the worst single request seen and a suggested client timeout.

API	p50 ms	p95 ms	p99 ms	max ms	p95/p50	failed	timeout
Open Library Search	1080	3672	8341	12379	3.4×	1/56	17 s
httpbin	852	2751	3809	4317	3.2×	0/56	8.5 s
Where the ISS at?	936	1466	1514	1522	1.6×	0/56	4.5 s
UK Carbon Intensity	341	1146	2025	2320	3.4×	0/56	4.5 s
Open-Meteo Forecast	696	1052	1375	1519	1.5×	16/56	3.5 s
Binance Ticker	359	790	886	917	2.2×	0/56	2.5 s
OpenStreetMap Nominatim	239	550	628	639	2.3×	0/56	2 s
Hacker News (Firebase)	391	508	557	575	1.3×	0/56	2 s
npm Registry	316	479	2705	5205	1.5×	0/56	5.5 s
GitHub REST	323	465	536	576	1.4×	0/56	2 s
Open-Meteo Geocoding	278	446	493	515	1.6×	0/56	2 s
USGS Earthquakes	147	409	1094	1638	2.8×	0/56	2.5 s
Datamuse Words	110	387	957	1628	3.5×	0/56	2 s
Zippopotam.us	211	373	471	475	1.8×	0/56	2 s
Wikipedia REST	220	355	387	395	1.6×	1/56	2 s
World Bank Indicators	163	307	350	361	1.9×	0/56	2 s
Coinbase Spot Price	169	292	318	326	1.7×	0/56	2 s
Nager.Date Holidays	94	240	321	342	2.6×	0/56	2 s
Frankfurter FX	96	230	3965	8390	2.4×	0/56	8 s
Dog CEO	151	225	263	268	1.5×	0/56	2 s
CoinGecko	80	222	297	304	2.8×	0/56	2 s
PokéAPI	102	221	243	257	2.2×	0/56	2 s
ExchangeRate-API (open)	87	190	269	310	2.2×	0/56	2 s
JSONPlaceholder	75	182	262	317	2.4×	0/56	2 s
PyPI JSON	106	175	244	278	1.7×	0/56	2 s

How the suggestion is calculated

Timeout = the larger of 2× the p99 or 3× the p95, rounded up to the next half second, never below 2 s. It covers the slow tail we actually observed without letting a hung request block you for a minute. Pair it with one retry after a short random delay: some APIs also fail briefly on the clock (see this finding). A p95/p50 ratio in red means the API is spiky: size your timeouts and your UI around p95, not the median.

Percentiles use successful checks only; the failed column counts all errors and timeouts. Small samples early in the week are noisy. Updated every 10 minutes. Machine-readable: /timeouts.json.