Agent Upstart

Practical notes on working with AI agents.

Playbooks · Coding tasks

Claude's watermark now covers older models — mostly

Written November 5, 2026

Mythos 5, Opus 4.7, and Sonnet 4.6 now carry Anthropic's text watermark. On Sonnet 4.6, one temperature setting can silently drop it.

What changed on November 3

Anthropic's November 3 entry in the Claude Platform release notes says text from Claude Mythos 5, Claude Opus 4.7, and Claude Sonnet 4.6 carries its text watermark. The support page's model table now ticks all three for text watermarks, on Anthropic's own surfaces and through cloud partners, with one footnote: rollout on Amazon Bedrock completes by November 10. Opus 4.6, named in the customer email that announced the first wave, is not in the November 3 notes — it moved to the second wave below.

The Sonnet 4.6 catch

On Sonnet 4.6, text from a request that sets temperature, top_p, or top_k to a non-default value "might not carry it." Anthropic's text watermark docs give one instruction: omit all three parameters from requests, including any a framework or client library sets for you. The other two models in the November 3 entry do not carry the caveat: Claude 4.7 and later models accept only the default sampling values.

The new API version closes the gap from the other side. In API version 2026-10-21, a request that sends temperature, top_p, or top_k returns a 400 error, even at the parameter's default value. A request that relies on a non-default value can stay on 2023-06-01 — and on Sonnet 4.6, that is the combination where the watermark may be missing. If a call lowered temperature for more consistent output, removing it changes how Sonnet 4.6 samples; no other setting replaces it.

The November 16 second wave is reported, not published

A second customer email, reported by The Decoder on November 4, says Opus 4.6, Opus 4.5, Sonnet 4.5, and Haiku 4.5 get the watermark from November 16, with Bedrock a few days behind. Anthropic has published nothing about that date: the release notes carry no November 16 entry, and the support page still leaves all four models blank for text watermarks. Treat November 16 as reported until an Anthropic page says it. The same report says all four models in the second wave share the sampling restriction — the watermark applies only while temperature, top_p, and top_k stay at their defaults — and that Opus 4.6 was delayed from November 3 without a stated reason.

What the mark is, and what it is not

The watermark is a version of Google DeepMind's SynthID-Text: a pattern in word choice, settled by a secret key, with nothing added to the text and no hidden characters. It adds no tokens, carries no information about the user or the organization, and does not change meaning, quality, or readability. It applies at the model level, so it covers the API, the Claude apps, Claude Code, and the cloud partners alike. Detection is in private preview for eligible organizations — regulators, media, fact-checkers, researchers, and similar groups — plus enterprises with their own compliance duty. Everyone else registers interest through a form.

A detected mark says only that Claude was likely involved with the content at some point. It cannot tell "Claude wrote this" from "Claude heavily edited this," and a missing mark does not mean the text is human-written — short passages, heavy edits, and exact output such as code carry less of a mark, or none. The driver is the EU AI Act: Anthropic applies the mark worldwide because it has no durable way to limit it by region, and models launched before August 2, 2026 are being brought in stages.

What to do

  1. List every call to Mythos 5, Opus 4.7, and Sonnet 4.6, and which platform serves it. For Bedrock traffic, treat marking as incomplete until November 10.
  2. For Sonnet 4.6, find every temperature, top_p, and top_k — including ones a framework sets — and remove the three where the mark matters.
  3. If a Sonnet 4.6 call needs a non-default temperature, keep that call on API version 2023-06-01 and accept that its output may carry no mark.
  4. For Opus 4.6, Opus 4.5, Sonnet 4.5, and Haiku 4.5, watch the support page's model table. Until Anthropic publishes the November 16 date, do not build a compliance claim on it.
  5. If detection matters, request access now rather than at audit time; most teams cannot check their own output today.

Limits

Related playbooks

Sources

Back to: Coding playbooks. Also: Move off Claude Sonnet 4.5 before it retires · Claude Code 2.1.318: sign-in expiry and fork effort.