Anthropic extends Claude text watermarking to older models including Mythos 5, Opus 4.7/4.6 and Sonnet 4.6, reportedly from November 3

Background and pain point

The EU AI Act's Article 50 transparency duties have applied since Aug 2, 2026, requiring generative AI providers to mark AI-generated content in a machine-readable way. Anthropic and about 190 other signatories joined the EU Code of Practice on Transparency of AI-Generated Content in July, and Anthropic explained Claude's text watermark on Aug 14. Until now only models launched after Aug 2 carried it; under the law's transition period, older models must follow. The official help center support table, updated Oct 28, now lists older models such as Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6. A customer email reported by media gives Nov 3 as the start date, with Amazon Bedrock possibly a few days later, but Anthropic's public pages do not state it. For people still using older models for resumes, copy or reports, output will become identifiable as likely Claude-assisted.

How it works

Anthropic says the watermark is a variant of SynthID-Text, published by Google DeepMind in Nature in 2024. It only changes the source of randomness used to choose among equally good words, adds no characters or hidden symbols, and uses no extra tokens. Detection uses Anthropic's key to estimate how likely a passage was partly written by Claude; it cannot prove human authorship or detect other AIs. Facts and code with a single correct token carry no watermark, light proofreading may leave too few changes to detect, and confidence grows with length.

Benchmarks

Anthropic has not published detection accuracy or false-positive rates. Citing the SynthID-Text paper, it says DeepMind's test on part of Gemini traffic found no statistically significant difference in thumbs-up and thumbs-down ratings, and its own internal tests found no effect on content, creativity or readability (vendor figures).

Getting started

API users need no code changes and pricing is unchanged. Three points matter: watermarks apply worldwide, not just in the EU; the detection API is in private preview, open by application through the help center form to regulators, law enforcement, media, fact-checkers, researchers and obligated enterprises; and older models currently carry text watermarks only on first-party surfaces, so check the support table for cloud-partner and C2PA coverage. If your product ships Claude-written text to end users, disclose AI involvement up front rather than relying on rewriting to strip the mark.