I didn’t notice that the Flash version had changed from 3.7 to 3.8 when I first asked Gemini a question earlier today. I wasn’t even aware of the announcement: Introducing Gemini 3.8 Flash and 3.8 Flash Cyber.

What I noticed, though, is that Gemini Flash 3.8 was slower than usual, even without Extended Thinking. It wasn’t particularly smart, either.

It wasn’t always stupid, but it struggled to clarify a Linux configuration issue. During a 17-question thread, it broke several times without responding (so I had to refresh the browser page). But the worst that happened was that, at some point, two strange things took place.

First, its answer to the 15th question (the 14th follow-up) was completely bogus and totally unrelated! It was Linux-related, but it mentioned IDEs, databases, lost data, and the swap, so it was almost certainly an answer to someone else’s question!

This is terribly frightening, if you ask me (and even if you don’t).

Then, it completely forgot everything discussed in the thread! I had to explain the context of my last question again, and it had to search the web one more time to rediscover what it already knew but had decided to forget.

In brief, it answered someone else’s question (obviously not the first answer in that person’s chat!), then it started afresh in a long thread as if nothing had happened!

How could I let Gemini 3.8 Flash touch my code? What if, instead of investigating a bug and observing the guardrails in AGENTS.md, it starts deleting whatever it feels like?

I’m afraid. I’m very afraid. You should, too, if you’re using Gemini Flash 3.8.

🤖

UPDATE: I wasn’t aware of that, but they say that ChatGPT, Grok, and Claude all went down at the same time, with someone saying, “Gemini also went down for me, I got Copilot to work but none of the others.” Wow. Oops.

According to The Reg, ChatGPT and Codex were restored at 8:17 am PT (15:17 UTC), and Claude was restored at 16:16 UTC. Grok’s outage was fixed at 17:07 UTC.