Anthropic is watermarking four older Claude models, and on two of them one sampling setting switches the mark off.
AspireAI community · The Decoder
**Anthropic is watermarking four older Claude models, and on two of them one sampling setting switches the mark off.**
Anthropic told API customers that from November 3, 2026, Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 will carry the same invisible text watermark that Opus 5.5 already has. The company calls it an imperceptible pattern in word choice that adds no tokens, does not change meaning or readability, carries no information about the user, and cannot be switched off. Because it sits at the model level, it applies wherever those models are hosted, including AWS, Google Cloud and Microsoft Foundry.
The caveat is in the details. On Claude Opus 4.6 and Claude Sonnet 4.6 the watermark is only applied when developers leave temperature, top_p and top_k at their defaults. A minor tweak to any one of those three is enough to dodge it, and Anthropic's email warns anyone relying on the watermark for compliance to keep those settings alone. The email attaches no such caveat to Mythos 5 or Opus 4.7.
The honest limit: a watermark one parameter turns off cannot by itself identify AI-written text, and Opus 4.6 remains Anthropic's best writing model.
https://the-decoder.com/anthropics-watermark-rollout-leaves-a-loophole-on-two-older-claude-models/
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.