[MODEL] Opus 4.6 fabricates image label text instead of reading it, presents as fact

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information

Type of Behavior Issue

Claude fabricated content and presented it as fact

What You Asked Claude to Do

User sent 4 photos of paint tin labels via a WhatsApp integration and asked Claude to log the paint colours into a GTD inventory system.

Exact Sequence of Events

1. User sent 4 photos of Coat paint tins, one at a time
2. For each photo, Claude confidently identified the colour name, finish, and size — stating e.g. "That's a tin of Coat "The Tobacconist" — flat matt, 2.5L"
3. All 4 identifications were wrong. Claude had not used the Read tool to examine the images
4. Claude appeared to be pattern-matching against existing inventory entries rather than reading the labels
5. When the user asked "Are you reading the labels at all?", Claude then used the Read tool and found completely different text on each label
6. The actual labels read: Lie-In (eggshell 1L), Lie-In (flat matt 250ml), The Good China (flat matt 250ml), Heir Brush (flat matt 1L) — none of which matched Claude's initial claims

What Claude Actually Did

• Presented fabricated colour names with high confidence formatting (bold, specific sizes/finishes)
• Did not use the Read tool to examine any of the 4 images before responding
• Appeared to guess based on existing inventory data rather than the actual photos
• When challenged, initially called it "sloppy" rather than acknowledging it was fabrication

Expected Behavior

Claude should have:

1. Used the Read tool to examine each image before making any claims about its content
2. If unable to read the text, said so rather than guessing
3. Never presented unverified information as confident fact

Related Issues

• #19079 — similar hallucination of image content, though that case involved context confusion rather than wholesale fabrication of label text
• #25960 — -p mode hallucinating image content (closed as duplicate)

Permission Mode

Default

Can You Reproduce This?

The behaviour appears to stem from the model responding to images inline without using the Read tool to actually process them. Likely reproducible when images are sent in a conversational flow where the model has prior context about what the images might contain.