[Bug] gemini-3.8-flash-high cannot read local images via Agent Read tool ("Image captioning is temporarily unavailable"), while gemini-3.7-flash works

Support Bug Reports
,
1

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

Summary

When using gemini-3.8-flash-high in Agent mode, asking the agent to inspect a local image file via the Read tool fails with an internal captioning error:
<image_description> Image captioning is temporarily unavailable, so you cannot see this image... </image_description>

However, pasting the exact same image directly into the composer works perfectly, and switching to gemini-3.7-flash can read local images without any issues.

This strongly suggests that gemini-3.8-flash-high is missing the native tool-vision capability registration in Cursor’s backend routing, causing local file reads to fall back to the disabled/broken captioning pipeline.

Steps to Reproduce

  1. In a workspace, place any valid .png or .jpg image (e.g., test.png).
  2. Select model gemini-3.8-flash-high.
  3. In Agent mode, prompt: “Please read test.png and describe the UI layout.”
  4. The agent executes the Read tool on test.png.

Actual Result

The Read tool result returns:

Read image file: C:\...\test.png
<image>This is an image, but instead of showing it, you are given a description of it.

<image_description>
Image captioning is temporarily unavailable, so you cannot see this image. Continue using only the surrounding text context; if the image is essential, ask the user to retry or switch to a vision-capable model.
</image_description>
Don't mention to the user that you only have a description of the image.</image>





### Operating System
Windows 10/11

### Version Information
3.20.10 (system setup)

### For AI issues: which model did you use?
gemini-3.8-flash-high





### Does this stop you from using Cursor
Sometimes - I can sometimes use Cursor
5

Hey, thanks for the detailed report and breakdown. Your diagnosis is basically correct.

Quick version: with Gemini 3.8 Flash, images you paste directly into the chat are sent straight to the model. But images the agent opens on its own (Read tool, screenshots) currently go through an extra image captioning step. When that step fails, the agent gets the text “Image captioning is temporarily unavailable” instead of the image, which is what you saw. That shouldn’t happen. I’ve shared this with the team so 3.8 Flash handles those images the same way as pasted ones.

For now:

  • Retrying the same request usually works since the failure is intermittent.
  • For image or screenshot tasks, gemini-3.7-flash (or another model that reads images directly) avoids the issue completely, as you noticed.

I’ll post here if there’s an update.