[BUG] Regression: screenshots pasted into chat are heavily downscaled and unreadable (v2.1.19)

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Until a recent Claude Code update (observed on 2026-04-15, running v2.1.19), screenshots pasted/dropped into the chat were passed to the model at (or near) their original resolution, and the model could clearly read text within the image — IDE code, log messages, error text, etc.

After the update, the same screenshots are downscaled very aggressively before being handed to the model. Fine detail (especially small on-screen text, like code in a Delphi IDE screenshot) is no longer legible. The model literally cannot read what it could read yesterday from visually identical input.

Reproduction

1. Take a full-window screenshot of a code editor / IDE at a normal desktop resolution (e.g. 2560×1400 PNG, ~230 KB).
2. Paste it into a Claude Code conversation and ask the model to read a specific line of code or text in the screenshot.
3. Compare to the behavior from ~a day ago with the same type of screenshot.

Example

• Original screenshot I pasted: 2560×1400, PNG, 237 KB — Delphi IDE with readable source code.
• What ended up visible to the model: a much smaller, heavily downscaled version where code text is blurred/unreadable.

I'm attaching both:

• the full-resolution original (what I actually pasted), and
• the downscaled version that reached the model (for comparison).

What Should Happen?

Screenshots pasted into the chat should be delivered to the model at high enough resolution that normal on-screen text (IDE code, logs, dialog text) remains legible — as was the case before this regression. If downscaling is necessary for size/bandwidth reasons, the threshold should keep text readable, matching the prior behavior.

Error Messages/Logs

Steps to Reproduce

1. Open a Claude Code chat session (v2.1.19 on Windows 10).
2. Take a full-screen/full-window screenshot at native resolution (e.g. 2560×1400 PNG, ~230 KB) of an IDE with small on-screen text — Delphi IDE code editor is a good stress test.
3. Paste the screenshot directly into the chat input (not via a URL, directly from clipboard or as a file drop).
4. Ask the model: "What does line N say in the code pane?" or "Read the text in the top-left watch list."
5. Observe: the model cannot read it — output is "I see a Delphi IDE but text is too small/blurry to read." Visually inspecting the image as it appears in the chat UI also shows a much smaller rendering than the original PNG.

Before today's update, the same workflow worked — the model could read the code clearly.

Claude Model

None

Is this a regression?

Yes, this worked in a previous version

Last Working Version

2.1.18 (or whichever version shipped immediately before 2.1.19 — behaved correctly up until today)

Claude Code Version

2.1.19 (Claude Code)

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Other

Additional Information

Shell: Git Bash (MSYS/MINGW) running on Windows 10 Home (10.0.19045). Terminal set to "Other" since Git Bash is not explicitly in the list.

Happy to provide a full-resolution sample PNG (e.g. the 2560×1400 Delphi IDE screenshot described in "Steps to Reproduce") on request — any screenshot of an IDE at native desktop resolution reliably reproduces the downscaling.