All Claude-generated text will carry embedded watermarks, including code and pull requests produced by Claude Code. The watermarks persist through copy-paste and survive some editing, though limitations exist.
New Claude models will embed invisible, in-text watermarks on all generated output as part of EU AI Act compliance. A text detection API is also coming so developers can verify Claude-generated content themselves. Other labs are doing similar things.
The watermarking layer operates below the model's awareness. Claude won't know it's embedding traces in its output. Google's SynthID is cited as a comparable reference point for how the technology works.
The embedded watermarking in Claude-generated text can be manually edited out, which the Anthropic team acknowledges. It's positioned as an imperfect but meaningful first step toward AI content attribution.
WebFetch currently uses Haiku for page summarization, which drew criticism for hallucination risk on fetched URLs. An Anthropic team member confirmed they're working to remove Haiku from WebFetch now that auto mode (which defaults to stronger models) is the default.
Following questions about whether WebFetch would drop summarized responses or swap to a different model, an Anthropic team member clarified the current direction is to try removing summarized responses entirely, though the work is still ongoing.
A server-side config change mistakenly displayed overages instead of subscription limits in the UI, though backend rate limits were always correctly applied. The 196 users who paid for Fable overages during the bug window have been fully refunded, and their accounts received bonus overage credits.
Some accounts stopped seeing the overall weekly usage limit, showing only the 5-hour and Fable limits. An Anthropic team member confirmed it's a bug affecting a small number of accounts and said it would be restored shortly.
A third-party benchmark claimed DeepSeek V4 Flash outperformed other models across four agent harnesses on 30 agentic tasks. An Anthropic team member pushed back, saying the results don't match internal daily evals, and asked for the methodology.
Git worktrees can accumulate fast and balloon disk usage with repeated node_modules copies. An Anthropic team member acknowledged the problem and floated building automatic stale-worktree cleanup directly into Claude Code.
More documentation on the watermarking rollout is available in Anthropic's help center, with additional details promised as the feature ships more broadly.
