Anthropic Brings Its Invisible Text Watermark to Four Older Claude Models, With Grok's API Next
Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 start carrying the mark on November 3; xAI turns on watermarking for seven Grok API models on November 7.
Published
Last updated / Fact-checked on
Anthropic will begin adding its invisible text watermark to four older Claude models on November 3: Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6. The Decoder reported the date from an email Anthropic sent API customers. xAI's Grok API follows on November 7, according to xAI's developer documentation. Both companies say they are doing this to comply with the EU AI Act.
Photo: viarami / Pixabay
Newer Claude models, launched on or after August 2, 2026, already carry the mark. The change closes a gap for developers still running older models.
Who is watermarking what, and when
| Provider | Models | Start date | Source |
|---|---|---|---|
| Anthropic | Claude models launched on or after Aug. 2, 2026 | At launch | Anthropic Help Center |
| Anthropic | Mythos 5, Opus 4.7, Opus 4.6, Sonnet 4.6 | Nov. 3, 2026 (Bedrock may take a few days more) | Customer email, per The Decoder |
| xAI | grok-4.7, 4.6, 4.5, 4.3 and three grok-4.20 variants | Nov. 7, 2026, on by default | xAI developer docs |
How the Claude watermark works
In an August explainer, Anthropic said the mark lives in Claude's word choices. When several words would fit equally well, the model normally picks one at random. With watermarking, a secret key and the preceding words decide that pick instead. The pattern is invisible to readers but can be checked by anyone holding the key.
Anthropic says the method is a version of Google DeepMind's SynthID-Text. It adds no hidden characters or extra tokens and costs nothing extra. It also carries no information about the user or organization.
The company is candid about the limits. Short passages, factual statements, code and light proofreading leave little room for the mark. A full rewrite removes it. A positive result shows only that Claude was "likely involved," not whether it wrote or edited the text.
Older models and a reported caveat
Anthropic's explainer said the EU law gives a transition period for models launched before August 2, 2026, and that watermarking for them would roll out "over the coming months." Its Help Center now lists the four models among those supported.
The Decoder reported that, per Anthropic's email, the mark on Opus 4.6 and Sonnet 4.6 applies only when the temperature, top_p and top_k settings stay at their defaults. Changing any of them may produce unmarked output. DailyMech has not seen the email itself.
Grok's API follows on November 7
xAI's documentation says text from seven Grok models will carry an invisible watermark starting November 7, on every API endpoint. It is on by default, and teams can opt out per model. A response header, x-watermark: true, shows whether a given output was marked. xAI describes detection as statistical, with longer text giving a stronger signal.
Who can check for the mark
Ordinary readers cannot test text for the Claude watermark yet. Anthropic says its detection API is in private preview for organizations eligible under EU law. Those include regulators, law enforcement, media, fact-checkers, independent researchers, educational groups and EU civil society organizations. Enterprises that must verify watermarking for their own compliance can also apply. Anthropic says it plans to expand access over time.
What to watch
- Remaining models. Anthropic has not given dates for Opus 4.5, Sonnet 4.5 and Haiku 4.5.
- Cloud partners. Timing on Google Cloud and Microsoft Foundry for the four older models.
- Grok opt-outs. How xAI squares per-model opt-outs with EU labeling duties.
- Detection access. Whether either company opens detection beyond the current groups.
For developers, older Claude versions no longer avoid the mark under default settings. Teams that mix Claude and Grok will see both changes land within the same week.
More coverage is on our Artificial Intelligence hub.
How we reported this
DailyMech read Anthropic's watermarking explainer and Help Center page and xAI's developer documentation directly. Anthropic's public pages do not give the November 3 date; that date and the default-settings condition come from The Decoder's report on an Anthropic customer email, which we did not see ourselves.
Published and fact-checked November 2, 2026.
Sources
- Anthropic, "How Claude's text watermark works," Aug. 14, 2026 (updated Sep. 1, 2026): https://www.anthropic.com/news/claude-text-watermark
- Claude Help Center, "How Claude marks AI-generated content," accessed Nov. 2, 2026: https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content
- The Decoder, "Anthropic-API-Nutzer können Wasserzeichen bei älteren Claude-Modellen mit einem Trick umgehen," accessed Nov. 2, 2026: https://the-decoder.de/anthropic-api-nutzer-koennen-wasserzeichen-bei-aelteren-claude-modellen-mit-einem-trick-umgehen/
- xAI developer docs, "Output Watermarking," accessed Nov. 2, 2026: https://docs.x.ai/developers/advanced-api-usage/watermarking
DailyMech used AI tools to help research and draft this article; it was fact-checked by our editors.