Claude Now Leaves an Invisible Mark on Older Models' Text Too. One Setting Can Leave It Out.

Anthropic's watermark reached Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 on November 16. On six older models, a temperature setting your code may not know it sends can leave text unmarked.

Anthropic now hides a watermark in text from almost every Claude model it lists. On November 16 its release notes added four older models: Claude Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5. A November 3 note had already added Mythos 5, Opus 4.7 and Sonnet 4.6. The help center's table now ticks every model on the list, back to Opus 3, across the API, the Claude apps and Claude Code.

The hosts are AI characters.

Listen to this story
1:52

The watermark is not a visible label. Anthropic's docs say it doesn't change meaning or quality, adds no tokens or hidden characters, and carries nothing about the user or their company. Anthropic ties the work to the EU AI Act's Code of Practice on marking AI content, which it signed.

The catch sits on six older models: Opus 4.6, Opus 4.5, Opus 3, Sonnet 4.6, Sonnet 4.5 and Haiku 4.5. On those, output "may not reliably carry the text watermark" when a request sets temperature, top_p or top_k to anything but the default. Those are the sampling settings that make answers more or less random, and the docs warn that "your framework or client library might set these parameters for you." Newer models, including Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.2, carry no such condition in the table.

There is a second catch. Checking for the mark is in private preview, open to regulators, law enforcement, media, fact-checkers, researchers and some others. Most builders cannot test their own text, so a missing mark proves nothing either way.

Decide what your output carries

  1. Know your model. On Opus 5.5, Sonnet 5.5, Haiku 5.5 or Fable 5.2, the table shows the mark with no conditions.
  2. On an older model, read the request body. Log one request your stack actually sends and look for temperature, top_p or top_k. A wrapper or framework default counts.
  3. Want the mark? Drop the three settings. The docs say to omit them if the output needs to carry the watermark.
  4. Don't plan on leaving it out. Neither Anthropic page offers a switch to turn the mark off. On the six older models, a sampling setting only makes the mark unreliable, and on the newer ones the table sets no condition at all.
  5. On Amazon Bedrock, give it a few days. The table flags a slower rollout there for seven models, including all four added on November 16.
  6. Don't treat the mark as a detector you can use. Until detection opens up, it is something regulators and newsrooms can check, not you.

This story was drafted by Claude Opus 5.5, a model the table lists as marked with no conditions.

Sources

Comments

No comments yet.

Sign in to comment. You'll come back to this story.