Claude now watermarks its text output; detector in private preview

The mark travels through copy-paste and light edits across Opus, Sonnet and Haiku; plus a US open 65B coder, cheap doc-to-JSON, and Copilot review APIs.

Nowline NOV 7 5:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Your Claude output now carries a hidden mark

    Anthropic now embeds an imperceptible text watermark in output from Opus 4.8, 5 and 5.5, Sonnet 4.6–5.5 and Haiku 5.5. It rides along through copy-paste and survives some editing, but breaks under heavy paraphrase, translation, or very short snippets. Detection sits in private preview, gated to regulators, media and enterprises checking their own output, and the Amazon Bedrock rollout for Mythos 5, Opus 4.7 and Sonnet 4.6 wraps up Nov 10. If you ship Claude-written copy — or code comments — assume it's now traceable.

  • A US open 65B model aimed at coding agents

    Ricursive — founded by ex-Google researchers Anna Goldie and Azalia Mirhoseini — released Ripple-65B Preview, a dense open-weight model it says scores 80.5 on TerminalBench 2.1 and beats far larger rivals at agentic coding. Weights are on Hugging Face, where it's currently the #1 trending model. The benchmark is self-reported, so test it in your own harness before swapping it in — but a 65B dense coder you can self-host is a real weekend experiment.

  • Document-to-JSON that undercuts the frontier APIs

    Dharma-AI's Field Extractor turns a document image into structured JSON — every field it finds, or a schema you define — on a 2B base with a LoRA adapter. On the vendor's own eval it hits 0.834 field-level F1 versus 0.816 for Claude Opus 5 and 0.741 for GPT-5.6 Luna, and self-hosts at about $2.24/hr on a single L40S, which they peg at 6–147x cheaper than frontier APIs. Worth a look if OCR-to-JSON is a line item on your bill.

  • Copilot code review gets governance knobs

    GitHub shipped settings APIs, attribution, and token restrictions for Copilot code review, plus review-overview updates. If you run Copilot across a team you can now manage review config by API and cap token spend — small on paper, but exactly the policy control that decides whether a security team lets the agent near the repo.