Every number needs a source.
Built for readers, developers and journalists who need to know what a reliability claim actually measures.
Official reports, retained over time.
We collect public provider status pages and feeds on a five-minute target. No social reports, user outage reports or synthetic API requests are mixed into the data. Every incident links to its official source and has a permanent Kingy URL.
Successful retrieval time is shown separately from the provider’s last update. A failed or unrecognized source becomes unavailable immediately. Old reports stay in the archive; a missing report never silently becomes resolved.
Coverage comes first.
Monthly uptime means reported incident-free time within continuously collected intervals. Two successful checks at most 15 minutes apart establish an interval. Longer gaps are unknown. We require 95% coverage of the elapsed UTC month or quarter before publishing a percentage.
Formula: 100 × (covered seconds − covered incident seconds) / covered seconds. Overlapping incidents are merged. Planned maintenance is excluded. Missing timing or ambiguous API scope prevents a qualified comparison.
A transparent recovery clock.
Mean time to recover is the arithmetic mean of known start-to-resolution durations for incidents resolved within the period. Where providers publish no impact window, we use first report to resolution and label that basis. Ongoing incidents and records with missing timestamps are excluded from the recovery average.
This is a reporting-based recovery measure. It can include time spent monitoring a fix, and is not guaranteed to equal actual customer downtime.
A qualified quarter, not a guess.
The quarter uses calendar boundaries in UTC and is provisional until it ends. Only explicitly identified API incidents are included. A winner across all six providers requires every source to meet the same coverage threshold. Ties share the lead.
Providers differ in incident disclosure, feature scope and geography. A partial regional incident counts as an interval, not a global outage. These figures are not request-weighted, independently measured availability or SLA compliance.
Source register
OpenAI ↗
OpenAI aggregate reporting; individual models, tiers and regions can differ.
Anthropic ↗
Public page and Atom feed. API scope requires an explicit API reference.
Google ↗
Google Cloud Vertex AI only. Gemini consumer app and AI Studio are separate; check aistudio.google.com/status for Gemini API.
xAI ↗
Official RSS incident reporting for Grok and API. A quiet feed means no active incidents reported, not a measured successful API request.
Mistral ↗
Public JSON endpoints. Blocked responses are treated as source unavailable; challenges are never bypassed.
DeepSeek ↗
Public page structured data and official Atom feed. Unknown formats fail closed.
Collection uses conditional requests, a named user agent and bounded response sizes. Access denials and rate limits trigger backoff; we do not bypass challenges, authenticate to private sources or crawl disallowed API paths on Claude or xAI. Source controls let the operator pause collection when terms or permissions change. Public access alone is not a license to republish complete incident narratives; Kingy retains normalized facts and links to the original.
Citing and reusing the data
Use a permanent incident or provider URL, cite the official source, include the retrieval date and specify service scope. Download the CSV for normalized incident facts and links. Historical imports are bounded by the feeds available at retrieval time; an empty archive or a zero incident count does not establish 100% availability.
Download incident CSV ↗