SpaceXSI documentation updated on October 31 says that supported Grok models will add an invisible watermark to API-generated text by default from November 7, 2026. The list includes Grok 4.7, 4.6, 4.5 and 4.3, as well as three variants of the 4.20 series. The provider says the setting applies across global and regional endpoints. As of November 4, the documented start date is still ahead.
What an application receives
The provider describes a method that slightly changes token selection as the model generates text. It adds no visible mark or extra hidden characters to the response. The change covers ordinary text responses, structured outputs and generated tool-call arguments. Applications do not have to add a request parameter. For the Responses and Chat Completions APIs, the response can carry an x-watermark HTTP header with the value true when the mark was applied.
The header describes one response as it travels through the API. If a team stores the text alone and forwards it to a different system, the HTTP header does not travel inside the text. An application that needs an internal audit record can preserve that value alongside its model and request identifier. Teams using streaming, caching or an intermediary should check whether that layer exposes the original response headers.
Detection on copied text is a separate operation
The provider describes statistical detection, with access granted through a separate application process. An uncertain result is possible, and the documentation says longer samples provide a stronger signal. This does not establish a universal authorship test. The documentation does not say what result to expect for an arbitrary short excerpt, translation or edit of the original response. These are different questions. One asks whether the API says it marked a particular generation. The other asks whether a piece of text later found outside that system contains a strong enough statistical signal.
One team setting
An administrator manages watermarking by model in the console. The choice applies to all API keys on the team. The documentation describes locked models and audit entries for setting changes. Opting out requires confirming that an exception applies to the particular use. SpaceXSI links the feature to European transparency requirements. The European Commission treats technical marking and disclosure to recipients as separate topics. The header can therefore serve as technical evidence, while public-facing disclosure and publication rules remain separate decisions for a given use.
For developers, the immediate check after the start date is straightforward. Send requests to the models you use in a test environment, record the metadata and verify that internal tooling carries it through. If multiple applications share a team account, confirm who owns the common setting before changing it.
