Live status of the major AI APIs from two angles: what the provider says on its own status page, and what a real HTTP probe to the API host returns from each of our four probe methods. The combination catches the common failure mode where the official status page lags real impact by 30+ minutes during major incidents.
ChatGPT, GPT-4o, GPT-4o-mini, the Realtime API, and the Assistants API. Powers Cursor, Windsurf, Notion AI, and most of the AI-flavored consumer apps shipped after November 2022.
Claude (Sonnet, Opus, Haiku), the Messages API, and the Claude.ai consumer surface. The other half of the frontier-model duopoly — every Cursor user with a Claude key, every Claude Code session, every coding-agent that ships in 2025.
Grok consumer chat, the api.x.ai inference endpoint, and the Grok integration shipped inside X / Twitter. The newest of the frontier-tier providers; status feed coverage is still maturing.
Mistral Large, Codestral, the Le Chat consumer app, and the open-weight family (Mistral 7B, Mixtral 8x7B). The European frontier lab — popular with EU-hosted apps that want a non-US inference vendor for data-residency reasons.
Ultra-low-latency LPU inference for Llama, Mixtral, Whisper, and DeepSeek-R1. The provider you pick when you need sub-100ms time-to-first-token — agent loops, voice assistants, real-time game NPCs.
Two probe layers, both free and unauthenticated. We poll each provider's official Atlassian Statuspage feed (e.g. status.openai.com/api/v2/status.json) every 5 minutes for the publisher's own classification. We separately run HTTP HEAD requests against the API host (api.openai.com) using 4 independent probe methods in parallel, that's the latency you see in each card. When the status page says green but our HTTP probes are timing out, that's the gap; we surface both signals so you can tell.
What this doesn't catch: model-level health (e.g. "GPT-4o is degraded but GPT-4o-mini is fine"), authenticated-API-only failures (we don't run paid requests), or rate-limit-induced 429s for specific accounts. Each provider's own status page handles those better than we could; we link to it from every card.