DigestAI news desk

Cut through the AI noise.

Policy & Regulation3 min read

Anthropic delays watermarking Claude Opus 4.6 to November 16

Anthropic announced a second round of watermark rollouts for its Claude models, delaying the addition of watermarks to Claude Opus 4.6 to November 16, 2026. The update affects four older models—Claude Opus 4.6, Claude Opus 4.5, Claude Sonnet 4.5, and Claude Haiku 4.5—with the watermark only activating at default sampling settings (temperature, topp, topk). This means developers can bypass the…

1 source

Key points

  • Anthropic delays Claude Opus 4.6 watermark to November 16, 2026, citing EU AI Act compliance
  • Watermark applies only at default sampling settings for four older models, allowing bypass via tweaks
  • Newer models (Mythos 5, Opus 4.7) lack this sampling restriction, but older models like Opus 4.6 remain bypassable

The company cites the EU AI Act’s December 2026 deadline as the reason for retrofitting older models. Anthropic’s watermark, based on Google DeepMind’s SynthID method, embeds an ‘imperceptible pattern’ in word choice without altering meaning, speed, or pricing. Critics, including blogger John Gruber, question whether the watermark degrades text quality by prioritizing statistical patterns over semantic accuracy. Meanwhile, legal concerns arise over potential detection in client documents where AI use is banned.

The story so far

4 episodes →
  1. Anthropic delays watermarking Claude Opus 4.6 to November 16this story
Full story from The Decoder · by Matthias BastianOpen source ↗

Anthropic's text watermark is easy to bypass on several older Claude models

The Decoder · 4 November 2026

Loading the full article…

This text was published by The Decoder and written by Matthias Bastian. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗

Topics · follow one to build your own front page

The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.

Comments

via GitHub Discussions

More in Policy & Regulation

All →

Related stories