Critical OCR Performance Regression: 2.5 Flash vs 3.1 Flash-Lite

Gemini API
1

Subject: Critical OCR Performance Regression: 2.5 Flash vs 3.1 Flash-Lite

Body: I am reporting a critical performance regression in the Gemini API. My production workflow, which has been reliably using ‘Gemini 2.5 Flash’, has broken down after the transition to ‘Gemini 3.1 Flash-Lite’.

  1. Context: I have been successfully using ‘Gemini 2.5 Flash’ for OCR tasks (CAPTCHA recognition). The accuracy was excellent.

  2. Issue: Since the update to ‘3.1 Flash-Lite’, the model is consistently producing incorrect results (hallucinations), making the API unusable for my application.

  3. Evidence:

    • Attached: Screenshots of my dashboard confirming the usage of ‘Gemini 2.5 Flash’.

    • Attached: 3 sample images that ‘2.5 Flash’ recognized perfectly but ‘3.1 Flash-Lite’ fails on.

  4. Demand:

    • Please investigate the performance difference between the 2.5 and 3.1 versions for OCR tasks.

    • I require an immediate solution to restore the performance level of ‘Gemini 2.5 Flash’ in the current environment.

Please escalate this to the engineering team responsible for model performance evaluation.

image
image1080×572 79.6 KB

image
image1080×1110 151 KB

2

I use Gemini API for OCR for production too and I noticed very bad regression of 2.5 Flash today. I believe I have found the problem regression between 2.5 models and the newer models. It is the ‘Media resolution’ setting. You need to set it to High to get good results. It appears that they may have defaulted it to a lower value for 2.5 Flash today.