Anthropic's watermark applies only to models after Aug 2

Anthropic will weave an imperceptible watermark into text from Claude models released on or after 2 August, applied at the model level so it travels through copy and paste and may survive some editing, and it will sign generated images and files with C2PA provenance metadata (The Verge). The date is not arbitrary. The EU AI Act's Article 50 transparency rules for AI-generated content have been enforceable since 2 August 2026 (Silicon Canals). The announcement is a compliance date wearing a product announcement's clothes.

The cutoff runs forward, not backward

Read the two dates against each other and the commitment is narrower than it sounds. The mark attaches to models released on or after 2 August, and on the record this ledger holds, Anthropic's two most recent model launches both landed before that line: Claude Opus 5 on 24 July 2026 (Anthropic) and Claude Sonnet 5 on 30 June 2026 (Anthropic).

So the marking begins with releases still ahead of the company rather than with the models doing today's work. That is the honest reading of a model-level mark: it is a property of how a model was trained, not a filter bolted onto an API, which is exactly what makes it hard to strip and exactly what makes it impossible to apply to something already shipped. Text produced by the models most people are using this week carries nothing. Whether that gap is months or weeks depends on a release calendar nobody outside the company has.

A rule with an uneven surface

Article 50 does not land as a single obligation. Visible labels, hidden machine-readable marks and editorial review each apply to different companies, different content and different moments (Silicon Canals). A lab that ships closed models through its own API can answer with a mark baked into the weights. That answer does not generalise. On the same day the watermark was reported, Nvidia released Nemotron 3.5 Lightning, a 30 billion-parameter mixture-of-experts open model, alongside an agentic router called NeMo Switchyard (SiliconANGLE). Two releases, one rule, and nothing that makes them comparable in how the rule reaches them.

Who can afford to go first

Building provenance into a model and signing every generated file is plumbing, and plumbing scales with balance sheets. Anthropic carries a $965B valuation on a $47B revenue run-rate as of May 2026 (Anthropic), and filed confidentially for an IPO on 1 June, three days after closing that round (TechCrunch). A company at that size and that close to a listing has reasons to be early on a European transparency rule that have little to do with the rule itself.

The useful question is not whether the mark works. It is what the first compliant model actually is, and when it arrives. Until then the rule is live and the marked text does not exist yet.

What the ledger holds

The record behind this note: Anthropic carries its funding rounds and release dates in one place, and the day-by-day feed is in the ledger. Every figure above links to the source it came from. Nothing here is estimated.