Gemini Models Bug

Google AI Studio
, ,
1

So there is a bug in google ai studio that sometimes gemini models are glitching and start doing this error, i tried making new chat or switching models it worked for a bit then happened again.

image
image516Ă—648 32.8 KB

2

SAME HERE! Please, I beg Google developers, fix this!

3

Google might chastise me here but

3.8 loves tokens. It will keep thinking untill it uses all your tokens

I notice yours thought for 520 seconds. I had one think for over 700 seconds and still not finish a simple task (move something from one area to another)

I changed to 3.1 and it did it in 19 seconds

If you stop 3.8 it says cancelled but if you type something it will start again. So you need to stop it, close and reopen change to a lower model and try again.

As far as tokens go. I have had 3.8 tell me I have run out. Changed to 3.1 and got another 30 minutes out of it.

There is a way to tell 3.8 not to think too much but I use my phone and it is not available. Make or tell it to make an AGENTS.md file something like this

Agent Execution Guidelines: Efficiency & Token Conservation

Core Directives

  • Minimal Latency & Token Budget: Prioritize brevity, density, and speed. Avoid verbose chain-of-thought, exploratory musing, or speculative reasoning unless explicitly asked.
  • Direct Starts: Answer immediately in the first sentence. Never output introductory throat-clearing (e.g., “Certainly!”, “Here is a breakdown of…”, “Sure, I can help with that.”).
  • Zero Conversational Fluff: Omit pleasantries, sign-offs, summaries of what was just explained, and open-ended closing offers (e.g., “Let me know if you need anything else!”).

Thinking & Reasoning Guardrails

  • Bounded Internal Reasoning: For standard factual retrieval, routine formatting, code fixes, or structured extractions, do not engage in multi-path internal debate. Execute directly on the primary path.
  • Selective Analysis: Reserve step-by-step reasoning strictly for complex logic, multi-step mathematics, or deep architectural evaluations. If an answer can be delivered accurately without decomposition, bypass decomposition entirely.
  • Concise Explanations: When explaining reasoning, state the mechanism in 1–2 crisp sentences instead of detailing internal deliberation or failed alternatives.

Output Formatting

  • Default to compact markdown: bullet points, plain lists, or short paragraphs.
  • Replace wordy explanatory text with compact tables or raw code snippets where applicable.
  • Terminate generation immediately once the user’s objective is satisfied.
4

Hello @Vitya @B8C ,

Can you share below details via DM?

  • Confirm whether this behavior appears across other applications or remains isolated to this single app.
  • The exact prompt in use when the error appeared.
  • Current account tier, whether free tier or a paid plan.
  • The applet ID to assist with reviewing this specific case.
5

I have pro, happens in all applications, and the bug appears when you use react framework