AI Press Box
Language: English
EspañolPortuguêsDeutschFrançaisItalianoУкраїнська

Four older Claude models now watermark their text, but a custom temperature may leave the mark out

Published 17 November 2026. Prices as recorded on that date.

Anthropic's Claude Platform release notes say in an entry dated 16 November 2026 that text generated by Claude Opus 4.6, Claude Opus 4.5, Claude Sonnet 4.5 and Claude Haiku 4.5 "carries Anthropic's text watermark on supported platforms". The same entry adds a condition: if a request sets temperature, top_p or top_k to a non-default value, "the text might not carry the watermark". Anthropic's Text watermark docs page says: "If you need your output to carry the watermark, omit temperature, top_p, and top_k".

On these models a developer's setting can decide whether the mark is there, and few outside Anthropic can check.

Which models carry the caveat, as of 17 November 2026

The Help Center page on how Claude marks AI-generated content says that for six models, outputs "may not reliably carry the text watermark".

ModelWatermark entry in the release notesNote
Claude Opus 4.616 November 2026Caveat stated in the entry
Claude Opus 4.516 November 2026Caveat stated in the entry
Claude Sonnet 4.516 November 2026Caveat stated in the entry
Claude Haiku 4.516 November 2026Caveat stated in the entry
Claude Sonnet 4.63 November 2026Same caveat, stated in that entry
Claude Opus 3No entryRetired from the API on 5 January 2026; researchers can request access

The same Help Center table lists the newer Claude Opus 5.5 and Claude Fable 5.2 as covered and marks no caveat against them, as of 17 November 2026.

Why the older models came later

Anthropic's post on the watermark, published on 14 August 2026 and updated on 1 September 2026, says it exists to comply with the EU AI Act and is "a version of the SynthID-Text approach published by Google DeepMind". Marking applies "wherever Claude is offered, worldwide", the Help Center says.

The same Anthropic post says the EU law gives a transition period to its models launched before 2 August 2026, and that Anthropic is working to watermark them too. The Help Center says Claude models launched in the EU on or after 2 August 2026 support machine-readable marking at launch. The 16 November 2026 note is the second such batch, after Claude Mythos 5, Claude Opus 4.7 and Claude Sonnet 4.6 on 3 November 2026.

Who can check

According to the Help Center, watermark detection for text is in private preview, open to eligible organisations such as regulators, law enforcement, media and researchers, and to select enterprises with their own duty to verify. The Help Center adds that a detected mark "provides a signal that content was processed by Claude, but is not fully conclusive".

What it means for developers and for the market

My reading: if your application calls Claude Opus 4.6, Opus 4.5, Sonnet 4.5 or Haiku 4.5 with a custom temperature, top_p or top_k, assume some of its text is unmarked, and you cannot confirm it unless you are in the preview. If your customers rely on Claude text being marked, dropping the three parameters is the lever you control. A model launched on or after 2 August 2026 is another route, though the only support for it is that the Help Center marks no caveat against Claude Opus 5.5 and Claude Fable 5.2.

Anthropic gives the EU AI Act as its reason, and the batches (3 November 2026, then 16 November 2026) suggest it works through older models a few at a time. For the market, a watermark that a sampling setting can weaken, checked by a detector few can use, gives a buyer a promise rather than something to test. In my view the detector decides whether watermarking matters to anyone outside the preview, and the Help Center says Anthropic plans to expand access "over time".

My forecast: by 17 May 2027 the Help Center will no longer describe text detection as a private preview, because Anthropic will have opened it more widely. If on that date the page still calls it a private preview, I was wrong.