Anthropic has expanded its invisible text watermarking to include Claude Sonnet 4.6 and Claude Opus 4.7, with the system now covering these models since November 3rd. This watermarking method, based on Google DeepMind's SynthID-Text, embeds marks within the model's word choices rather than adding characters. However, a caveat exists for Sonnet 4.6: if the temperature, top_p, or top_k parameters are set to non-default values, the watermark may not be present, creating a potential gap for users relying on this feature. AI
IMPACT This update impacts developers by potentially altering how AI-generated text can be identified, with specific implications for Sonnet 4.6 users who rely on non-default sampling parameters.
RANK_REASON This item details an update to an existing feature (watermarking) for specific model versions, rather than a new model release or significant research breakthrough.
- Amazon Bedrock
- Anthropic
- AWS
- Claude API
- Claude Code
- Claude Mythos 5
- Claude Opus 4.7
- Claude Sonnet 4.6
- Google Cloud
- Google DeepMind
- Microsoft Foundry
- SynthID-Text
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →