Loading live status…
Notes on the services we watch — what breaks, what a provider's status page does and does not tell you, and how to work out whether an outage is theirs or yours.
LinkedIn had trouble on 13 August 2026. Its status page filed the incident already resolved, with zero duration, no components attached, and an impact level that leaves the indicator green — eighteen hours after the fact. Here is what that does to every monitor pointed at it, including ours.
LinkedInOne million reports in a few weeks, and a 40% drop in views for content LinkedIn classifies as AI slop. If your distribution runs through LinkedIn, that is an availability event — and it is the kind no status page will ever report.
LinkedInClaude reports on six surfaces — claude.ai, the API, the Console, Claude Code, Cowork and Gov — and its incidents are usually scoped to one model rather than to all of them. How to read that page, and the five failures that look like an outage and are not.
Coinbase publishes 174 status components — one per asset, per payment rail, per surface. Most "Coinbase is down" moments are one of those 174, and a good share of them are a blockchain being slow rather than Coinbase being broken.
ElevenLabs reports on eleven components, and they split into two halves that fail for completely different reasons: the model path that generates audio, and the telephony path that carries it. Which half you are standing on decides what a green status page means for you.
Firebase does not publish a component grid like most providers — it publishes an incident log, per product and often per region. That difference changes what "green" means, and it is why a Firebase outage can be entirely real and entirely invisible to you.
FirebaseGitHub publishes eleven components that fail independently, so "GitHub is down" is almost never true of all of GitHub. How to read its status page properly, the five failures that impersonate an outage, and what still works while it is genuinely broken.
GitHubThe classic interview question, answered for a web that has moved on. HSTS before the first packet, HTTPS records that skip the TCP handshake, QUIC instead of TCP, TLS 1.3 in one round trip — and a map of which stage each kind of outage actually lives in.
Binary framing, multiplexing, HPACK, and the two headline features that were later removed. What HTTP/2 actually changed, where it loses to HTTP/1.1, and how to tell which protocol a site is really serving.
Ten mechanisms that take down large sites — config pushes, BGP withdrawals, expired certificates, split-brain databases — and, for each one, what it looks like from outside while it is happening.
Every uptime number is an artifact of when somebody looked, who did the looking, and what happened when the looking failed. Four bugs from building isuptime, and what each one says about reading anyone's status page.
In two weeks Cloudflare gave AI agents payment rails, Anthropic made Claude Code act without asking, and Vercel replaced agent credentials with scoped tokens. Three companies independently concluded the bottleneck is no longer intelligence — it is permission.
Vercel
CloudflareOpenAI updated GPT-5.6 Sol for Plus and Pro users, made GPT-5.6 Luna the default for Free, and added a slider that controls how hard the model thinks. Here is what OpenAI actually claims, and what it means if you build on the API.
OpenAIVercel Ship 2026 delivered eve, Vercel Connect, Vercel Services, Dockerfile support and AI SDK 7 — and the August changelog kept going. A read on what Vercel actually built this year and what it says about where the platform is heading.
VercelSkip the reading — check the live dashboard instead.
Open the live dashboard