I hope you’re doing well. I am writing to report a persistent issue I’ve been experiencing with the Gemini2.5-pro-exp model on my paid plan.
Issue Description:
Whenever I request a response, the model repeatedly halts generation after only a small portion is produced. This happens almost immediately—after just a few lines—and then again and again. As a result, each attempt consumes one of my paid-plan question allocations without delivering a complete answer.
Impact:
Wasted Queries: Every interruption forces me to re-submit or continue the request, using up extra queries.
Workflow Disruption: I cannot rely on the model to produce full, uninterrupted answers.
Ongoing Problem: This has been occurring for several days and shows no sign of improvement.
Steps to Reproduce:
Send a prompt to Gemini2.5-pro-exp (e.g., “Explain the theory of relativity in simple terms”).
Observe that the generation stops after only a few sentences.
Attempt to continue; it stops again shortly thereafter.
Expected Behavior:
The model should continue generating until it delivers a complete response or reaches a natural stopping point per the prompt.
Could you please investigate this issue and let me know if there is a workaround or a fix in progress? I rely on Gemini2.5-pro-exp for my daily work, and these interruptions are significantly affecting my productivity.
Thank you for your prompt attention to this matter. I appreciate any assistance you can provide.
Same situation, I tried creating a topic but it got hidden for some reason.
Few days ago I noticed a massive downgrade in the gemini 2.5 pro performance. When tried to continue development of my project which gemini worked perfectly on before, but now its completely unusable.. It doesnt try to understand the project files anymore and just hallucinates everything and gets stuck on a loop to fix things that never ends up working. It also has started adding random zod schemas which it hasn’t done before.
Using latest version too, tried using different versions but nothing changes.
Tried with any other model like sonnet 3.7 and works fine.
I assume they changed their prompts to be more “vibe coding friendly” meaning having it add zod schemas, proper logging libraries like “logger” which it does now automatically when adding changes so and it messed up everything It doesn’t even follow my cursor rules anymore but as soon as I switch to sonnet 3.7 then that model works just like its supposed to.
I am also experiencing this and it is quite frustrating that the premium credits are consumed for nothing. I would gladly live with this annoying bug when I could turn off fast responses and use slow ones instead. It feels like getting ripped off when fast responses are consumed for this behavior. For really small or minor changes I would gladly use slow responses and only switch to fast ones for bigger tasks. I am basically out of fast responses in the first week of every month.
Hey, we’re working with Google to fix these issues as we are unable to detect them when they occur, unfortunately, but we are confident these issues should be entirely ironed out soon!
If you have lost a significant amount of requests on this, please feel free to email us at [email protected]
We have a significant volume of queries, but one of the team will get back to you!