Claude Watermarks Are Live, But Not on the Models You're Using
Anthropic's Claude watermarks went live August 2, 2026, but Claude Sonnet 5 and Opus 5 both predate the cutoff and aren't marked yet. Here's what's actually covered, the backlash, and the removal tool already out.
Contents6
Claude watermarks are live, but not everywhere you'd expect
Anthropic's Claude watermarks went live on August 2, 2026: invisible marks woven into text output, plus signed C2PA provenance metadata on supported files like PNG, JPG, and SVG. The trigger is Article 50 of the EU AI Act's transparency code, but Anthropic applied the rule globally instead of geo-fencing it to Europe, so the change hit every Claude user at once, not just EU accounts.
Here's the detail most of the coverage glossed over: the rule only applies to Claude models launched on or after August 2, 2026. Claude Sonnet 5 shipped June 30. Claude Opus 5 shipped July 24. Both are the models most people and most paid teams are actually running right now, and both fall into what Anthropic calls a "transition" group that isn't watermarked yet, with no public date for when that changes.
What actually gets marked, and what doesn't
| Output type | Marking method | Survives editing? |
|---|---|---|
| Text (marked models only) | Invisible statistical watermark woven into token choices | Persists through copy-paste; degrades with heavy rewriting |
| Supported image files (PNG, JPG, SVG) | Signed C2PA provenance metadata | Stripped if the file is re-exported or flattened |
| Claude Sonnet 5 / Opus 5 output | Nothing yet | N/A (pre-dates the cutoff) |
| Code output | Same text watermark as prose | Coders report concern it alters token patterns |
The mark applies across the API, the Claude apps, Claude Code, Cowork, and Tag, and it follows you onto AWS Bedrock, Google Cloud, and Microsoft Foundry too. Anthropic says it's working on retrofitting the older, more heavily used models during the Act's transition period, but as of this writing there's no committed timeline.
The reaction hasn't been friendly
A TechCrunch piece on the rollout led with users angry that the watermark could get them caught using Claude at work or in school, which tells you where a chunk of the userbase's actual objection lives. It isn't really about privacy in the abstract, it's about a tool that used to be invisible now leaving a trail.
On Hacker News, a thread ostensibly about why Opus 5 "feels worse" to work with lately floated the watermarking rollout as a possible cause of the stylistic drift some developers reported. That's a theory, not a finding, and the thread itself was split on whether it holds up. Worth knowing about, not worth treating as settled.
Developers using Claude for code raised a more concrete complaint: nobody's fully sure whether biasing token selection to embed a watermark has any measurable effect on code quality, and until Anthropic says more about the mechanism, some teams are choosing to assume the worst.
Watermark removal already exists
An open-source project called watermarks-remover added support for stripping Claude, OpenAI, and Gemini provenance marks within days of the Anthropic rollout. It's straightforward to verify whether it worked on the C2PA file metadata, since that's just a metadata field you can inspect. It's a different story for the statistical text watermark: there's no reliable way to confirm removal actually happened, because the whole point of a statistical watermark is that you can't visually tell it's there in the first place. That cuts both ways: you can't prove you removed it, but a checker can't easily prove it's still present either once the text has been substantially rewritten.
What this means if you use Claude today
If your workflow runs on Sonnet 5 or Opus 5, nothing changes yet. If you're on an older, already-marked model or you're testing new releases as they ship, assume future models carry the watermark from day one and design around that rather than treating it as a bridge you'll cross later. For a broader look at what changed with Anthropic's current flagship, see our [Claude Opus 5 pricing and benchmarks breakdown](https://questloops.com/blog/claude-opus-5-explained-pricing-benchmarks-and-where-it-actually-wins), and for the regulation driving all of this, our [Article 50 explainer](https://questloops.com/blog/the-eu-ai-act-s-transparency-rules-just-went-live-what-article-50-actually-requires) covers what the transparency code actually requires. The [Claude tool page](https://questloops.com/tools/claude) has the current plan and pricing breakdown.
Quick take
Anthropic didn't quietly ship a global watermark, it shipped a *partial* one, and the gap between "watermarking is live" and "the models you use are watermarked" is the part worth remembering. Check back before you assume either way.
