Hot AI News

Anthropic now watermarks text from four older Claude models, too

Anthropic says output from Claude Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 now carries its watermark, though changing sampling settings may stop the mark from appearing.

Anthropic now watermarks text from four older Claude models, too
Image: Anthropic

Anthropic has extended its text watermark to four older models: Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5. The mark applies on supported platforms, according to the Claude Platform release notes dated November 16, 2026.

Four older models join the watermark list

The entry covers four models that predate the newest Claude generation. Anthropic’s notes point to a support article, “How Claude marks AI-generated content”, for the list of platforms where the mark applies. The release notes themselves do not spell that list out.

Anthropic now watermarks text from four older Claude models, too
Image: Anthropic

There is one stated limit. A developer who changes temperature, top_p or top_k away from its default on any of these four models cannot count on marking. In Anthropic’s wording, the output “might not carry the watermark”. A separate Text watermark page gives the details.

The same release notes show the rollout has been going on for months. An entry dated November 3 says output from Claude Mythos 5, Claude Opus 4.7 and Claude Sonnet 4.6 is marked as well, with the same caveat for Sonnet 4.6 when sampling parameters are changed. An entry from September 1 covers Claude Fable 5.1 and Claude Mythos 5.1, and says that supported image, video and audio files produced through the code execution tool carry C2PA Content Credentials when retrieved through the Files API. Marking needs no changes to requests or response handling, according to that entry.

Advertisements

What a developer needs to check

For most API users nothing changes in code. The practical effect is for teams that tune sampling. Anyone running these four models with custom temperature, top_p or top_k values cannot assume their output is marked. Anthropic says only that the mark “might not” be present, so the notes do not say how often it is lost or under which settings.

The notes also do not say how a third party can detect the mark, and they give no accuracy figures. These are Anthropic’s own systems, and nothing in the entry is independently verified.

A compliance step that is arriving model by model

Reporting on Anthropic’s August announcement ties the marking to Article 50(2) of the EU AI Act, which requires providers of generative AI to make output detectable as artificially generated. According to New Atlas, Anthropic signed the EU Code of Practice on Transparency of AI-Generated Content, one of about 190 signatories, and applies the marking worldwide because it does not yet have a durable way to limit it by region.

The same report says models launched on or after August 2, 2026 support marking at launch, while older models get it over the coming months under a transition period the EU law allows. The November 16 entry fits that pattern: these four are older models being brought in line. A third-party summary on ChatGPT Is Eating the World notes that Anthropic did not state the length of that transition period.

New Atlas also describes the method as a version of Google DeepMind’s SynthID-Text. A secret key and the preceding text nudge the choice among plausible next words, leaving a statistical pattern that a detector holding the key can check. Anthropic says, per that report, that this adds no extra tokens, does not reduce quality and does not identify the user. For a reader, that means light editing probably leaves the mark, while a full rewrite removes it, and a detected mark shows Claude was likely involved, not that a human did not write the text. The report adds that a detection API is planned but not finalized.

The sampling caveat fits that design. A statistical bias in word choice is harder to apply when a developer has already reshaped the probability distribution, which would explain why those three parameters matter. Anthropic does not give that reason in the notes, so it remains an inference.

Advertisements

Who is affected and what is still unclear

Anyone using the Claude API, or the apps on the supported platforms, with these four models gets marked text by default. The timing matters for one of them: the same release notes announce retirement of Claude Sonnet 4.5 on the Claude API on November 30, 2026, with Claude Sonnet 5.5 as the recommended replacement.

The entry leaves several questions open: the full platform list, how detection will work for the public, and what share of sampled output carries the mark. Until Anthropic ships its detection tooling, the mark is something only the company and its partners can reliably read.

Sign Up for the Daily AI Pulse

One email a day. All the stories that matter.

Leo MartinsAI Tools, Prompts & Practical AI Expert

Leo Martins is the AI Tools & Practical AI Expert at Aiholics, focused on helping readers use artificial intelligence to improve productivity, creativity and everyday work.

Leave a Reply

Your email address will not be published. Required fields are marked *

You may also like