AI news · September 6, 2026

When Claude, ChatGPT, Grok, and Gemini went dark on the same day

A strong September model week — then half your stack 500’d. FOMO, token anxiety, and a spare-tire plan for vibe coders.

Read time: ~9 minutes

September started loud. New models every other day. Then Thursday morning, a chunk of the vibe-coding stack just… stopped. Not a bad prompt. Not a rate limit. Dark.

If you build with Cursor, Claude Code, ChatGPT, or Grok, you felt it in your gut before you checked Downdetector: that weird FOMO where the project is right there, the deadline is real, and the only thing between you and “ship” is someone else’s status page.

This post is the Thursday outage, the model flood that opened September, and a practical spare-tire plan so the next blip does not freeze your whole week.

What went down (3 Sep 2026)

Around the same window on Thursday 3 Sep, the big chat / coding surfaces many vibe coders live in hit elevated errors or full outages:

• OpenAI ChatGPT + Codex — elevated errors; chat, logins, uploads, voice, search, deep research, and image tools were hit for a stretch. OpenAI described a routing error starting roughly 7:43 a.m. PT.
• Anthropic Claude — partial outage across Claude.ai, Claude Code, Claude Cowork, and the Claude API (Anthropic: “infrastructure issue”). Staff noted recovery around 12:15 p.m. ET.
• xAI Grok — overloaded / unavailable across app, web, and X for a stretch; “try again or pick a different model.”
• Google Gemini — Reddit threads and Downdetector reports spiked the same morning (“can’t connect / generation failed”). Google did not publish a matching status-page incident or confirm a consumer outage. Same-day coverage still counted the Downdetector jump (≈23 → 412 reports) plus a brief StatusGator “likely outage” window on the Gemini API (~10:45–11:15 a.m. ET). Some outlets argue Gemini mostly rode it out on Google Cloud while Azure-heavy stacks blinked; builders who hit dead sessions that morning still felt a fourth outage. I’m counting Gemini as the fourth — community + monitor signal, not a Google announcement.

Cause: still muddy in public reporting. Same-day coverage lined up Claude/ChatGPT/Grok without a confirmed shared root. Separate write-ups finger shared cloud concentration (Azure East US ingress theories) for some stacks — which also explains why Gemini press is split. Until vendors publish postmortems, label shared-cloud theories as hypothesis, not fact.

Either way, the user lesson does not need a postmortem: your daily driver is not infrastructure you control.

Why it hits vibe coders harder

Vibe coding trains a loop: describe → agent runs → review → next prompt. When the model is gone, the loop does not degrade gracefully into “write a function by hand.” It stalls.

That stall has a name people have been using all year: token anxiety — the feeling that something could be running right now, so rest feels like falling behind. An outage is token anxiety with the power cut.

Common Thursday feelings (honest ones):

• Research tab open, agent mid-refactor, nothing to do but refresh status
• Deadline work stuck behind one vendor’s API
• “I’ll just switch models” — until you realize your whole project context lives in one chat / one vendor
• Guilt that peers on another stack might still be shipping

None of that means you are weak. It means the unit of work moved from your laptop to their uptime.

September started strong — then the lights flickered

Same week (and the days around it) the launch feed did not slow down — including GPT-6 Astra landing on the same calendar day as the OpenAI outage:

What a September start. Labs will keep shipping models. Thursday was the reminder that shipping models ≠ shipping 100% accessible service. Capability without uptime is a demo. Uptime without a spare model is a single point of failure on your calendar.

Practical spare-tire plan (do this once)

1. Name a backup before you need it. Cursor model picker, Claude ↔ GPT ↔ Gemini ↔ Grok ↔ OpenRouter open models. Pick one “if primary dies” default per project.
2. Keep at least one free/cheap BYOK key warm. Walkthrough: Free AI API keys for vibe coding (2026) — Google AI Studio, Groq, OpenRouter, etc.
3. Do not trap context in one vendor chat. Export or mirror the brief (goal, constraints, file list) in a local NOTES.md so a cold start on another model is five minutes, not an hour of re-explaining.
4. Split agent risk. Long cloud agent on Cursor Origin / Claude Code is fine — just know which jobs can wait vs which need a second provider.
5. When everything is dark: ship the human parts. Spec the next ticket, clean the README, write the QA prefill, review the Graveyard list. Packaging work from How I create Apify Actors… does not need a frontier model.
6. Watch status, don’t doomscroll. OpenAI / Anthropic / Google / xAI status pages beat Twitter theory threads.

Honest cliffs

• Multi-provider is not free — you pay in setup and in “which model has my style.”
• OpenRouter and free tiers have their own outages and quotas.
• Shared-cloud concentration means “three logos” can still be one failure domain. Diversity of brand ≠ diversity of datacenter.
• Token anxiety does not go away when uptime returns. Build a stop rule: agents off after N hours, or when the next prompt is just dopamine.

FAQ

Were all four — Claude, ChatGPT, Grok, and Gemini — fully down?

Claude, ChatGPT/Codex, and Grok have confirmed vendor/status trails. Gemini is the contested fourth: Google did not announce it, but Reddit groups, Downdetector spikes, and a short StatusGator “likely outage” window lined up the same morning. Some coverage says Gemini mostly stayed up on Google Cloud. This post still counts four because that is how many builders experienced it.

Is vibe coding “addiction” just a meme?

The funny version is a meme. The useful version is token anxiety: infinite capacity creates guilt about rest. An outage makes the dependency visible.

What’s the fastest backup if Cursor’s default model dies?

Switch model in-picker first. If the whole vendor is dark, OpenRouter + a free/cheap key from Part 1, or Google AI Studio if Gemini is healthy for you.

Bottom line

Labs will keep shipping models. Thursday’s ask: ship accessible service too — including regional rollout. Keep a spare tire warm before the next status-page red.