On this page
xAI’s official developer documentation for Output Watermarking—the HTML page carries Last updated: October 31, 2026—states that text responses from supported Grok models will include an invisible watermark starting November 7, 2026. Read against 2026-11-05 in Asia/Shanghai, that start date sits about two days out. The page does not name a clock time or a time zone for the switch. Default watermarking is not treated here as already on across the board.
The stated aim is compliance with the EU AI Act. The docs refer to the product as SpaceXSI in several places—phrases such as “SpaceXSI watermarks…” and “SpaceXSI Console”—while the API endpoint remains https://api.x.ai, including regional endpoints. DigIntoAI treats that as the same xAI / Grok API line, not as a second company or an acquisition story. A nearby note on the brand sits at xAI Opens the SpaceX Super Intelligence Partner Network.
Watermarking is described as enabled on every SpaceXSI API endpoint, including the global endpoint https://api.x.ai and regional endpoints. Control is per model for the whole team. There is no request parameter to send; an integration works the same way with or without the watermark.
What follows tracks xAI’s documentation as self-reported. DigIntoAI has not actually run the interface and has not checked a live x-watermark header—the checking steps are for reference only.
Related DigIntoAI notes on other vendors’ marks: Claude Text Watermarks Now Cover Mythos 5, Opus 4.7, and Sonnet 4.6—Detection Still Private and OpenAI textGrain: Invisible Watermarks in the EU. A separate detector write-up, DetectorScope Explains Its AI-Text Calls. Why That Still Isn't Proof of Cheating, is a third-party auditable AI-text tool—it is not xAI’s official Detection API.
Starting November 7: the model list
Starting November 7, 2026, responses from the following models include a watermark. The feature is on by default, and Team admins can opt out per model:
grok-4.7grok-4.6grok-4.5grok-4.3grok-4.20-reasoninggrok-4.20-non-reasoninggrok-4.20-multi-agent
In short: Grok 4.7／4.6／4.5／4.3 and the 4.20 series (reasoning／non-reasoning／multi-agent). Models outside that list are not written here as receiving the same default watermark on the same schedule.
Which outputs carry the mark
Watermarking applies to text output from text generation requests, for each model that has watermarking turned on for the team.
It applies to every text generation API, including the Responses API and the Chat Completions API. It covers the generated text received in the response: ordinary text, structured outputs, and tool call arguments.
The documentation page used here does not extend that rule to image, video, or voice products, so coverage stays on API text generation.
How to read x-watermark
Whether a given response was watermarked shows up on the x-watermark header of the Responses API or Chat Completions API response.
x-watermark: truemeans the output for that request was generated with the watermark.- Any other value, or no header at all, means it was not.
No request-side parameter is involved. Apart from the x-watermark header, the response format is unchanged. The docs’ FAQ puts the same point in one line: calling code does not need to change for watermarking itself.
Official examples all inspect that header and all use grok-4.7 as the sample model: the AI SDK (@ai-sdk/xai plus generateText), Python with the OpenAI SDK and with_raw_response, JavaScript with the OpenAI SDK and .withResponse(), and curl -i against https://api.x.ai/v1/responses. Working details belong on xAI’s Output Watermarking documentation; the operational takeaway is to look at the response header. DigIntoAI has not replayed those examples against a live key.
What the watermark is (briefly)
The watermark is not visible. It is applied while the model generates text: it slightly shapes which tokens the model picks, and it adds no characters, markup, or hidden text to the response.
Detection is statistical. The system scores a piece of text against the watermark and reports whether it is watermarked, not watermarked, or uncertain. Longer text gives a stronger signal.
Other vendors have shipped their own invisible text marks; the Claude and OpenAI notes linked above cover those products. The mechanics are not repeated at length here—the useful part for Grok API callers is the model list, the header, and who can turn the feature off.
Who can opt out—and who cannot
Only Team admins can turn watermarking off for some or all models in the Console. The setting lives on the Team settings page, on the Model output watermarking row, which shows On (all models), Off (all models), or Custom. Only team admins see that row.
The documented path: open Team settings, find the row, select the settings icon next to it, then clear the checkbox for each model to opt out. Select all turns every model on or off at once. After Save, turning a model off requires a further Opt Out confirmation.
Changes apply to new requests within about 5 minutes. After a model is opted out, its new requests stop including the watermark; after it is opted back in, new requests start including it again.
The opt-out confirmation, in the docs’ wording, asks the admin to confirm that an exception under applicable law applies to the use, and that the team is solely responsible for any marking, disclosure, or other legal requirement. That is a documentation condition, not a legal forecast.
Some models may be locked on for a given team. Locked models appear in the dialog but cannot be turned off from the Console (they show without a checkbox). Questions about a locked model go to the team admin, or to support@x.ai.
The setting belongs to the team, so it applies to every API key on that team. The choice is made separately for each model. It is not a personal one-click toggle available to every key holder.
Every opt-in and opt-out is recorded in the team’s audit log. Team admins can view that log in the Console and search by event ID, description, or user, then narrow results with a time filter.
A separate FAQ line states, in the docs’ own wording, that watermarking doesn't change response quality or speed. That sentence is repeated here as documentation language, not as a lab measurement.
Detection API: application required
Access to watermark detection is granted by application. The documentation points to a Google Form for requesting the Detection API. That path is not a public endpoint open to every caller, and it is not the same thing as DetectorScope Explains Its AI-Text Calls. Why That Still Isn't Proof of Cheating. DetectorScope is a third-party write-up about an auditable AI-text detector; applying for xAI’s official Detection API still goes through the form linked from the Output Watermarking docs.
No detection accuracy rates or enforcement outcomes are added here. Detection remains a statistical three-way call (watermarked／not／uncertain).
Not tested yet
DigIntoAI has not placed a live Grok API call against the listed models for this feature, has not inspected a production x-watermark header, and has not submitted a Detection API application. Everything above follows xAI’s Output Watermarking documentation as self-reported on a page last updated October 31, 2026.
Quick facts
- When. Starting November 7, 2026; about two days after 2026-11-05 (Asia/Shanghai). No clock time or time zone in the docs.
- Docs stamp. HTML Last updated: October 31, 2026.
- Why. EU AI Act compliance, per the docs.
- Brand. xAI／Grok API; docs also say SpaceXSI; endpoint
https://api.x.ai(global and regional). - Models.
grok-4.7／grok-4.6／grok-4.5／grok-4.3／grok-4.20-reasoning／grok-4.20-non-reasoning／grok-4.20-multi-agent. - APIs. Responses API and Chat Completions API.
- Coverage. Ordinary text, structured outputs, and tool call arguments.
- Header.
x-watermark: true= watermarked; any other value or no header = not watermarked. - Requests. No parameter to send; response format unchanged apart from that header.
- Opt-out. Team admins only; Console Team settings, Model output watermarking; Select all available; about 5 minutes; team-wide across all API keys; audit log; some models may be locked on.
- Detection. Application required (Google Form); statistical watermarked／not／uncertain; not DetectorScope and not a public free-for-all.
- Quality／speed. Docs: watermarking doesn't change response quality or speed.
- Not tested. DigIntoAI has not actually run the interface or checked a live header.
- Not covered. Grok chat apps and Grok on X sit outside this API text write-up.
In short
xAI’s Output Watermarking docs (last updated October 31, 2026) say that starting November 7, 2026, listed Grok 4.7／4.6／4.5／4.3 and 4.20-series models will add an invisible text watermark by default on Grok API text generation—Responses and Chat Completions, including ordinary text, structured outputs, and tool call arguments—to comply with the EU AI Act. Whether a given reply was marked shows up on the x-watermark header (true versus any other value or a missing header). Only Team admins can opt out per model in Console, with changes landing in about five minutes, an audit-log trail, and the possibility that some models stay locked on. Detection access is by application. DigIntoAI has not actually run the interface; for reference only, from the official documentation.