From the Anthropic edition of October 31, 2026
Anthropic tells API customers the Claude text watermark is extending to four older models, but on Opus 4.6 and Sonnet 4.6 it applies only at default sampling settings
Anthropic (reported by The Decoder) · AI model provider; all Claude API and cloud-marketplace customers · Global (EU AI Act driver); applies wherever Claude is offered
- Publisher
- The Decoder
- Original publication
- October 31, 2026
- Source retrieved
- 2026-11-01
What happened
Per The Decoder's account of an Anthropic customer email, the Claude text watermark is being added to Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6. On Opus 4.6 and Sonnet 4.6 the watermark applies only when temperature, top_p and top_k are left at default values, so changing any one of them removes it; the email tells customers relying on the watermark for compliance to keep defaults. Detection remains in private preview.
Why it matters
Relevant to any agency, school or university that builds on the Claude API (directly or through Amazon Bedrock or Google Vertex) and expects to rely on provenance marking for records, public communications or academic-integrity processes. Why now: the change is announced this week and the parameter caveat means application settings, not the vendor alone, decide whether output is marked. Limited relevance for staff using only the Claude apps.
Evidence and measured results
Independent trade reporting of a vendor email, with partial direct quotes. No measurements of detection reliability.
Limitations and uncertainty
Based on a reporter's reading of a vendor email; confirm against the vendor model table and terms. Watermark quality effects are disputed by commentators, not measured here.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-11-01; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- Agencies and institutions may assume all Claude output carries a provenance watermark when coverage depends on model and sampling settings.
- Stakeholders
- Records management officers
- Public information officers
- Academic integrity offices
- Application development leads
- Discovery questions
- Which Claude models and versions do your applications call, and through which platform?
- Do any applications change temperature, top_p or top_k?
- Does any policy or contract assume AI-generated text is watermarked?
- Value hypothesis
- Interpretation: clarifying watermark coverage early can prevent policies and procedures that overstate what provenance marking proves.
- Potential engagement
- A short provenance-readiness review that maps Claude models and settings in use against stated watermark coverage.
- Unsupported claims
- That all Claude output is watermarked
- That a watermark proves authorship or misconduct
- That detection is available to U.S. agencies or schools today
Pre-sales engineering
Role takeaway
- Fit
- Relevant to any custom application on the Claude API, Bedrock or Vertex that uses Opus 4.6 or Sonnet 4.6.
- Architecture and integration
- Sampling parameters are set per request; a configuration registry or gateway can enforce defaults where marking is required.
- Prerequisites
- Inventory of model versions per application
- Record of sampling parameters in use
- Decision on which workloads need provenance marking
- Constraints
- Marking on Opus 4.6 and Sonnet 4.6 depends on default sampling
- Bedrock rollout may lag by days
- Detection API is private preview
- Security
- The watermark reportedly carries no user or organization data; it does not create a new personal-data flow and cannot identify an individual.
- Proof of value
- Confirm model-by-model coverage against the vendor table
- Test whether gateway-enforced defaults hold across applications
- Verify Bedrock or Vertex coverage dates before relying on marking
Delivery
Role takeaway
- Work
- Document provenance-marking configuration for each Claude application and bring parameter changes under change control.
- Dependencies
- Vendor coverage documentation
- Application owners
- Change-control process
- Ownership
- Application owners, with the records or academic-integrity office as policy owner.
- Skills and adoption
- Developers need to understand that sampling settings affect marking; policy owners need plain-language limits.
- Governance checkpoints
- Coverage review before go-live
- Change-control check on sampling-parameter changes
- Annual policy review of provenance claims
- Acceptance criteria
- Proposal: 100% of Claude applications have recorded model version and sampling settings
- Proposal: no policy text claims universal watermarking
- Proposal: parameter changes on affected models require documented approval
- Risks
- Policy overstates provenance
- Silent parameter changes remove marking
- Detection access remains unavailable to U.S. public bodies
Implementation considerations
Lighthouse Advisory interpretation across the operating dimensions a public-sector buyer must settle before this evidence becomes a design. Each note answers the question under its heading for this specific source.
Architecture and integration
What must connect, and where does the AI sit in the workflow?
Application teams that tune sampling parameters on Opus 4.6 or Sonnet 4.6 will produce unmarked output. Platform teams should inventory model versions and sampling settings per application, and note that Bedrock rollout may lag first-party surfaces by days.
Governance
Who approves, reviews and stays accountable for outcomes?
Any policy that assumes Claude text is always watermarked is inaccurate for some models and settings. Governance documents should state which models and configurations are covered.
Security and privacy
What data, permissions and controls need testing?
Per the reported email, the watermark carries no user, organization or conversation information, so it does not create a new personal-data exposure; it also cannot attribute text to a person.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
Staff and faculty guidance should not imply watermark detection can prove authorship; training should explain coverage gaps.
Procurement
What should contracts, pricing and exit terms secure?
Buyers who need provenance marking for compliance should ask the vendor to document model-by-model coverage and default-setting dependencies in writing, including for cloud-marketplace delivery.
Operating model
Which teams own the service once it runs?
Changing a sampling parameter becomes a governance-relevant change on affected models; change control should flag it.
Publication history
- 2026-10-31Anthropic · Issue 037 resources
Stable resource ID: decoder-claude-watermark-older-models-sampling-default-loophole-20261031