ChatGPT, Claude and Grok Went Down Together. What That Says About How We Use Them.

TL;DR: On 3 September, ChatGPT, Claude and Grok all degraded within the same two-hour window. Gemini mostly stayed up, because it runs on different infrastructure. Two AI vendors sitting on the same cloud is one dependency with two logos. And the deeper issue is not downtime, it is how much work we hand to a single model that was never the right tool for most of it.

01The Most-Searched Thing About ChatGPT Was Whether It Worked

Last week's search data has ChatGPT at half a million queries. Almost none of them wanted the product.

"is chatgpt down". "chatgpt status". "downdetector". "ai down". "why is chatgpt not working".

And inside the same search breakdown: claude ai, gemini, grok. People checking the alternatives.

The alternatives were down too.

02What Happened

OpenAI opened an incident around 10:58 UTC on 3 September. Nineteen degraded components across ChatGPT and Codex. Conversations, login, file uploads, voice, image generation. Downdetector logged over 74,000 reports.

Anthropic flagged a disruption in the same window. So did Grok. Everything was back by early afternoon Pacific time.

Gemini mostly kept working. Its peak was around 500 reports, against ChatGPT's 74,000.

The difference was infrastructure. Gemini runs on Google's own cloud. The others shared Azure.

03Two Vendors, One Cloud

That is the part worth remembering.

Plenty of teams already have a second AI provider. It feels like a backup. It goes on the architecture diagram as a backup.

On 3 September it was not a backup, because both providers were standing on the same floor.

If you rely on AI for anything that matters, it is worth knowing which cloud each of your providers runs on. Most people have never asked, because a model API looks like a service rather than software running on somebody else's computer.

Worth knowing too: OpenAI's own published figure for ChatGPT availability between June and September was 99.62%. That sounds excellent. It is about 2.7 hours of downtime a month.

04The Bigger Question

Outages are the visible problem. Here is the quieter one.

Ask ChatGPT or Claude to convert a document, read text off a scanned receipt, check whether an email address is real, or pull a table out of a PDF, and it will try. Sometimes it does a decent job. Often it produces something that looks right and is wrong in a way you will not notice until it matters.

We tested that. On a clean, machine-rendered invoice, a standard OCR tool read every word correctly and got two of seven monetary values right, dropping decimal points while reporting 93 to 96% confidence. The full test is here.

So when your AI assistant goes down, two things stop working. The things it was genuinely good at. And the things it was doing badly that you had not noticed yet.

05Give the Model Tools Instead of Tasks

There is a better shape for this, and it has nothing to do with redundancy.

A language model is very good at understanding what you want and deciding what to do. It is not good at converting a video, running OCR on a photo, checking a VAT number against a live registry, or telling you today's exchange rate. Those need real machinery or a live lookup, and no amount of reasoning produces them.

MCP is the standard that lets an assistant use real tools instead of improvising. You connect it once, and Claude or ChatGPT can find the right capability and call it.

Every endpoint in the ApyHub catalog works this way. That changes three things.

Fewer tokens. Asking a model to write a document converter costs thousands of tokens across attempts, errors and retries. Calling one that exists costs a few hundred and returns the actual file.

Better output. A purpose-built OCR endpoint is not guessing. Neither is a VAT lookup that queries the actual registry.

More control. You can see which tool was called and what it returned. When something is wrong, you know where.

The catalog covers file conversion, OCR and document extraction, data validation, domain and company checks and more. One key, and your assistant finds what it needs.

There is a resilience benefit too, though it is the smaller point. Work that runs through a specialised endpoint does not stop because a model provider is having a bad morning.

See the MCP setup | Get a free API key

06Three Things Worth Doing

Find out which cloud your providers run on. If your primary and your backup share one, you do not have a backup.

Notice how much you are asking the model to do. Anything involving a real file, a live lookup, or a fact about right now is probably better handled by something built for it.

Try connecting one set of tools. Add the catalog to Claude or ChatGPT and give it something it would normally fumble. Watch which checks it decides to run. That takes about a minute and it is more convincing than reading about it.

07Conclusion

Three assistants went down together because they shared a floor. That is worth knowing.

But the more useful takeaway is that a lot of what we hand to a general-purpose model was never a model problem. Converting files, reading scans, verifying data. Give it the right tools and you get better answers, cheaper, from a system you can actually inspect.

Browse the catalog | Start free, no card

08FAQ

Why did ChatGPT, Claude and Grok all go down at the same time?

No provider published a root cause. Reporting attributed the overlap to shared Azure infrastructure. Gemini stayed largely operational because it runs on Google's own cloud, with a peak of around 500 Downdetector reports against ChatGPT's 74,000.

How long was the outage?

OpenAI opened the incident around 10:58 UTC on 3 September, and full restoration across ChatGPT, Claude and Grok was confirmed by early afternoon Pacific time. Roughly a two-hour window.

Is using two AI providers enough of a backup?

Only if they run on different infrastructure. Check which cloud and which region each one uses before treating a second vendor as redundancy.

What is MCP?

MCP, or Model Context Protocol, is a standard that lets AI assistants connect to real tools and use them. Instead of a model improvising a solution, it finds a capability built for the job and calls it. It works with Claude, ChatGPT and most current AI tools.

How does MCP reduce token usage?

Asking a model to build something costs tokens on every attempt, including the failed ones, and those stay in the conversation. Calling an existing endpoint costs a fraction of that and returns a real result on the first try.

What should I not ask a language model to do?

Anything requiring a live lookup, such as today's exchange rate or whether an email inbox exists. Anything requiring real processing, such as video conversion or OCR on a photograph. It will attempt all of them and the results tend to look plausible while being wrong at the edges.

09Further Reading

OCR in Python - the decimal point test, and where document reading goes wrong
How AI Agents Use APIs - what MCP is and how assistants find tools
API Fundamentals - if none of this vocabulary is familiar yet
Top 5 OpenRouter Alternatives - the model routing layer and who owns it now

10Sources

11About ApyHub

ApyHub is a curated catalog of ready-to-use APIs for developers, teams and AI assistants. Every endpoint is MCP-ready, so Claude, ChatGPT and other AI tools can find and use them directly: file conversion, OCR and extraction, data validation, domain intelligence and more across 20 categories.

One key covers everything, with usage pooled across the catalog rather than locked to individual services. Every service carries machine-readable certification covering data handling, retention and standards alignment including GDPR, SOC 2 and ISO 27001.

ApyHub is EU-based and runs entirely on EU infrastructure, which keeps data residency simple for teams with GDPR obligations. The catalog holds 450+ services and 1,500+ endpoints, with new APIs added continuously. The free tier requires no credit card.

ChatGPT, Claude and Grok Went Down Together. What That Says About How We Use Them.