Skip to content
sobolewski.proAI today, every eveningSubscribe

Once a day. Unsubscribe with one click. Privacy policy

Claude 4.5 watermark arrives, Opus 4.6 too: mind the temperature

The Claude 4.5 watermark went live on November 16, along with Opus 4.6. Set your own temperature, though, and the text may go unmarked. Here is who that affects.

3 min read

This article in other languages:Polski

In this article
  1. The Claude 4.5 watermark, and Opus 4.6
  2. What are these four models still missing?
  3. What should we watch out for?
  4. FAQ

The Claude 4.5 watermark is live: as of November 16, text generated by Claude Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 carries Anthropic’s invisible watermark on supported platforms. The news appeared in the Claude Release Notes, together with a caveat that this time applies to all four models.

The Claude 4.5 watermark, and Opus 4.6

This is the second batch in two weeks. On November 3 the watermark came to Mythos 5, Opus 4.7 and Sonnet 4.6, which I covered in my post on the Claude watermark (it also explains how the mechanism works).

After today’s change, the table on the help page on content marking lists 18 models. Put it next to the list on the model deprecations page and it turns out that every model with “Active” status now has the watermark. Of the models still running, only Claude Mythos Preview, which is being retired, is left out of the table.

Interestingly, this batch includes Sonnet 4.5, a model that has been “Deprecated” since September 30 and will be switched off on November 30. So it got its watermark two weeks before the end (better late than never ;)).

What are these four models still missing?

Newer models have three columns ticked in the table: the watermark on Anthropic’s own surfaces (Claude Platform, the Claude apps, Claude Code, Claude Cowork and Claude Tag), the watermark with cloud partners (AWS, Google Cloud, Microsoft Foundry) and C2PA metadata in files.

For Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5, only the first of these is ticked as of November 16. The Release Notes do say “supported platforms”, not all of them, so if we use these models through a cloud provider, it’s better not to count on the watermark for now.

What should we watch out for?

First, the sampling parameters. If a request to any of these four models sets temperature, top_p or top_k to a non-default value, the text might not carry the watermark. In the previous batch, only Sonnet 4.6 had that caveat in the Release Notes. Today the watermark documentation names six models: Opus 4.6, Sonnet 4.6, Opus 4.5, Sonnet 4.5, Haiku 4.5 and the retired Opus 3 (available on request).

The advice in the documentation hasn’t changed: if we rely on the watermark, we leave out all three parameters and check that a framework or library isn’t setting them for us.

Second, the API version. The exception applies to the older models for a simple reason, which we can find in the migration guide for API version 2026-10-21: Claude 4.7 and later accept only the default values of these parameters, and on the new API version a request with temperature, top_p or top_k ends in a 400 error, even at the default value. So the gap stays open only where an older model meets the older API version (2023-06-01).

Third, what a missing watermark means. According to the help page, it doesn’t prove that a model didn’t write the text: the cause may be heavy editing, a very short text or an unsupported platform. We can’t check it ourselves anyway, because text watermark detection is in private preview for selected organizations. I described a different approach to the same question, a cryptographic proof of which model gave the answer, when writing about FlashProve.

FAQ

Which Claude models have the watermark?

According to the table on the help page (as of November 16, 2026):

  1. Fable 5.2, 5.1 and 5,
  2. Mythos 5.2, 5.1 and 5,
  3. Opus 5.5, 5, 4.8, 4.7, 4.6 and 4.5,
  4. Sonnet 5.5, 5, 4.6 and 4.5,
  5. Haiku 5.5 and 4.5.

Can the Claude 4.5 watermark be turned off?

The help page doesn’t mention such an option. A custom temperature, top_p or top_k only means the text “might” not carry the watermark, so we can’t rely on it either way.

Do I need to change anything in my API requests?

No. According to the documentation, the watermark adds no tokens or hidden characters and requires no changes to requests or to response handling.

  • AI
  • Anthropic
  • Claude
  • AI Act

AI today, every evening

The day's biggest AI news and new posts from the blog, every evening. Short, with sources.

Once a day. Unsubscribe with one click. Privacy policy