2.5 Flash Model blocking images based on content being prohibited, 2.0 Model is fine, and Open AI's competing model is fine

Gemini API
,
1

We use the Flash 2.5 model and have been using it to process imagery (receipts in this case).

It’s been fine for Flash 2.0 (and still is), but the same images are being blocked by the new 2.5 model.

Same images, two different outcomes of 2.0 (fine) vs 2.5 (not fine).

I’ve tried adjusting safety settings, I’ve tried with thinking on and off.

Still the same issue, if there’s any skin in the photo (i.e. hand, finger) it prohibits content responses.

OpenAI’s gpt-4.1-mini is fine also, I will be forced to change to that if you cannot resolve this quickly.

2

Hi @DrewB, Thanks for reporting this issue. If possible could you please the sample images where the 2.5 model fails and works fine with the 2.0 flash model to reproduce the issue. Thank You.

4

Hi @DrewB, May I know the prompt or any Preprocessing steps you are using. Thank You.

5

I’m not keen on offering up what is our IP around system prompts and others which competitors could copy/follow. Can I email them directly or send a private message?

6

Hi @DrewB, You can DM me those details. Thank You.

7

I’m unsure how to send a direct message on here sorry, I checked your profile and there’s no link.

9

Hi @DrewB, I have tried to extract the details from the given image with Gemini 2.5 flash with and without passing the structured out the execution went well and I did not face any error as you said image are blocked. I have shared the gist link in the DM you can check that. Thank You.