SpaceXSI's Grok documentation describes default watermarking for supported text models starting November 7, 2026. The Output Watermarking page says the feature requires no new request parameter and applies across the global and regional API endpoints.
What the watermark changes
The mechanism slightly shapes token selection during generation. It does not append visible characters, markup or hidden text to the response. The documentation includes ordinary text, structured output and tool-call arguments within the supported text-generation coverage.
The listed model family includes Grok 4.7 and several earlier variants. The announced date is still ahead of this article's preparation, so the documentation should not be read as evidence that every response is already watermarked today.
How an integration checks a request
The Responses and Chat Completions APIs expose an x-watermark response header. A value of true indicates that the particular output was generated with the watermark. A missing header or another value does not mean the same thing.
Keeping that request-level distinction is more informative than assuming a team setting establishes the state of every historical response.
Detection has limits
The publisher describes statistical detection that can return watermarked, not watermarked or uncertain. Longer text provides a stronger signal, and detection API access is granted through an application process.
That differs from a publicly available cryptographic proof that a specific model produced a passage. The page does not establish that all copied or edited text can always be identified reliably.
Team configuration
Watermarking is configured per model for a team and applies to its API keys. Some models may be locked on, and setting changes are logged. Developers should consult the current documentation for their access and obligations rather than assuming the configuration or detection service is identical for every account.