What happened on Oct 10, 2025
On the morning of October 10th (JST), ChatGPT suddenly stopped opening external URLs via open_url.
No error, no warning—just silence. Search-style answers still worked, which is why most users didn’t notice.
But the ability to open and read a specific page vanished.
Why many users didn’t notice
web.search() continued to return summaries. For everyday queries, that feels “good enough.”
Only when you ask ChatGPT to open a specific URL and extract a concrete field (like the page title) does the failure appear.
My stack didn’t stop
Ten days before this event, I had finished a Dify + Firecrawl setup.
So when open_url stopped, my system still worked.
- Firecrawl crawls and normalizes web data
- Dify stores that data as Knowledge
- ChatGPT only generates answers based on my Knowledge, not live URLs
Because of that, I could keep “reading the world” through my own architecture.
Why this matters
This isn’t a random bug — it’s a structural shift.
AI systems are being redefined to use controlled, licensed retrieval paths.
From “freely reading the web” to “reading through designed architectures.”
Design, not usage
That’s why I say: Stop just using AI. Start designing it.
Because the tools will keep changing,
but your own system will always adapt.
🌍 English readers: You can read this article in your language using your browser’s translate feature.
This post shares first-hand insights from Japan about AI and ChatGPT updates.
🔗 Japanese Original (KenjiNext47AI)
🔗 Series Hub — AI Learning Summary
コメント