Anthropic watermarks older Claude models, with a gap

If you publish text Claude wrote, assume it carries a mark you can't see. Now that goes for older Claude models too.

Anthropic's help page on how Claude marks AI-generated content, updated this week, says it is "adding watermarks to outputs from models released before August 2, 2026." Models launched in the EU on or after that date carry the mark from launch.

The mark lives in the words

Anthropic says supported Claude models weave "an imperceptible watermark directly into the text itself." It says the mark doesn't change the meaning, quality or readability of a response, and that it may survive copying, pasting and some editing.

Files are marked differently. Supported image files, such as PNG and JPEG, get signed Content Credentials (C2PA) metadata.

The help page lists where text marking applies. That covers Anthropic's own surfaces (the Claude Platform API, the Claude apps, Claude Code, Claude Cowork and Claude Tag) and Claude on AWS, Google Cloud and Microsoft Foundry.

Anthropic says it has signed the EU AI Act's Article 50(2) Code of Practice on transparency of AI-generated content.

Detection is narrow for now. Anthropic says its watermark detection is in private preview for groups such as regulators, law enforcement, media, fact-checkers, researchers and educational organizations, plus enterprises with their own verification duties. A free Claude Content Checker verifies the C2PA credentials on files.

The fine print is in the sampling settings

The Decoder reports that Anthropic emailed API customers about extending the mark to Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6. According to The Decoder, the email says customers cannot turn the mark off, and the rollout on Amazon Bedrock may take a few extra days.

Here's the thing the email reportedly flagged. On Opus 4.6 and Sonnet 4.6, The Decoder says, the mark applies only when temperature, top_p and top_k stay at their defaults. Change one and the text goes out unmarked.

Anthropic's help page does not mention that condition. TopFive could not open the email itself.

What it means for you

Two kinds of readers should care, and they want opposite things.

If you need the mark, for an EU compliance duty or a client who asked, leave the sampling settings alone on those two older models. The Decoder reports Anthropic told developers exactly that.

If you'd rather your drafts not carry a mark, know that the choice isn't yours. The Decoder reports the mark is applied at the model level, so it follows Claude onto every cloud that serves it.

Either way, Anthropic's own page is blunt about the limits. It says "a detected mark provides a signal that content was processed by Claude, but is not fully conclusive," since Claude may only have proofread or translated the text. It adds that a missing mark doesn't prove a human wrote it, because heavy editing, translation and short passages can all strip it.

So treat a detector result as one piece of evidence when it lands on your desk.

TopFive covered the same tradeoff when Google opened its SynthID Detector to everyone. The AI watermarking bargain hasn't changed. A mark is easy to add and easy to lose.

What to watch

Watch who gets into Anthropic's detection preview, and whether it opens wider. A mark only matters to the people who can read it.

Then check your own settings. On two models, your defaults decide whether the mark is there.