Claude now watermarks all its text output — detector API coming
Every passage Claude writes now carries a hidden SynthID signal; code and edited text mostly slip through. Plus Cursor's 3x-faster agents and new Grok tools.
Copy markdown
Different randomness, invisible signature
Claude now uses Google DeepMind's SynthID-Text: an encryption key plus preceding context steer word choices instead of arbitrary randomness, leaving a detectable pattern in every passage it writes — translations included — with no extra tokens and no added cost.
The catch: code and edited prose slip through
The signal needs room to vary, so it's weak or absent in code (exact outputs required), short factual passages, and heavily proofread text. If you ship Claude-written prose, assume it's marked; if you ship Claude-written code, mostly not.
Global today, detector still 'being worked out'
The rollout applies worldwide immediately for current models; pre-Aug-2 models get it over the coming months. Verification needs Anthropic's key and the detection API isn't live yet. Image and file outputs get C2PA provenance metadata instead of a text watermark.
Cursor Cloud Agents boot up to 3x faster
Cursor's new 'builds' pre-clone your repo and pre-install dependencies, so cloud agents start up to 3x faster and resume from the last good build when an environment breaks. A Builds tab adds status and logs; auto-enabled for everyone by Aug 17.
Grok video and 4.6 land in builder stacks
Vercel's AI SDK added xAI's Grok Imagine Video 1.5 with native 1080p output and reference-audio-to-video. Separately, Grok 4.6 — tuned for long-running agents — is live in Cursor at $2 per million input tokens with 2x included usage this launch week.