Grok API Watermark Goes Default on Seven Models Today

Starting November 7, 2026, text from seven Grok API models carries an invisible watermark by default. SpaceXSI (formerly xAI) documented the change on its Output Watermarking developer page, citing EU AI Act compliance. There is no new request parameter. The response body looks the same. The signal lives in token selection, plus an x-watermark response header you can read if your client keeps headers.

That is a product change for anyone shipping with the Grok API today: agents, SaaS features, search tools and internal copilots. It is not a consumer-app announcement, and the docs do not say whether Grok on X or the grok.com chat UI use the same mark.

What SpaceXSI turned on

The watermark applies to text generation from these models, on by default, with a per-model opt-out for team admins:

• grok-4.7
• grok-4.6
• grok-4.5
• grok-4.3
• grok-4.20-reasoning, grok-4.20-non-reasoning and grok-4.20-multi-agent

It covers the Responses API and the Chat Completions API, on every SpaceXSI endpoint including the global https://api.x.ai host and regional endpoints. The mark touches ordinary text, structured outputs and tool-call arguments—the strings your agent may pass into tools—not only free prose.

Metarix's Grok tool page and the new SpaceXSI company hub are the places we keep model and policy context for this stack.

How you tell a response was watermarked

Read the x-watermark header on the Responses or Chat Completions reply. A value of true means that request's output was generated with the watermark. Any other value, or a missing header, means it was not.

SpaceXSI says the watermark itself is not visible. During generation the model slightly shapes which tokens it picks. No characters, markup or hidden text are added to the string you see. Detection on the provider side is statistical: longer samples give a stronger signal, and the detector can return watermarked, not watermarked or uncertain. Access to that detection API is by application, not a public self-serve tool in the docs.

The company states watermarking does not change response quality or speed. That claim is from the vendor FAQ; Metarix has not independently measured latency or quality deltas.

Opt-out is a compliance decision, not a debug flag

Team admins can turn watermarking off for some or all models in the SpaceXSI Console under team settings. The choice applies to every API key on the team. Changes hit new requests within about five minutes and are written to the team's audit log.

Opting out is not free of process. The Console asks you to confirm that an exception under applicable law covers your use, and that you alone are responsible for any marking, disclosure or other legal requirement. Some models may be locked on for a team and cannot be unchecked. Locked models appear in the dialog without a checkbox; SpaceXSI points those questions to the team admin or [email protected].

That framing matters more than the toggle. Leaving the default on keeps SpaceXSI's mark on your completions. Turning it off moves marking and disclosure duty onto your org. For product and legal teams, that is the decision—not whether a string looks different in the UI.

How this sits next to the EU AI Act timeline

SpaceXSI explicitly ties the feature to EU AI Act compliance. Metarix already covered the EU AI Act delay: which obligations were postponed and which still apply. Watermarking and provenance for AI-generated content sit in that same policy stack. Other labs are shipping different technical answers—cryptographic provenance in some cases, statistical marks in others. Builders should treat "default on" as the durable setting unless compliance signs off on an opt-out.

What builders should do this week

• Preserve response headers. Proxies, gateways and SDKs that strip headers will hide x-watermark. If you need an audit trail, log the header with request IDs.
• Inventory the seven models. Any path still pinned to grok-4.7, 4.6, 4.5, 4.3 or the 4.20 variants is in scope the moment the default flips.
• Decide opt-out with legal, not only engineering. The Console language assigns marking responsibility to you if you disable the mark.
• Watch tool-call arguments. Watermarking includes tool-call argument text. If you hash or cache those strings for idempotency, expect the same semantic content with a different statistical signature across runs—behavior you already get from sampling, now with an official mark on top.
• Do not assume consumer Grok is identical. The docs scope the change to API text generation. Consumer surfaces are out of scope of this write-up until SpaceXSI says otherwise.

Verdict

This is a real, dated product change: default-on invisible watermarking for seven Grok API models from November 7, 2026, with a header to inspect and a Console opt-out that shifts legal responsibility. Code that only looks at the JSON body needs no rewrite. Teams that care about provenance, EU exposure or gateway header passthrough should update runbooks today. Treat vendor "no quality or speed impact" as unverified until you measure your own workloads.

Sources

• SpaceXSI docs: Output Watermarking (models, header, opt-out, FAQ; accessed Nov 7, 2026)
• Metarix: EU AI Act Delay — what's postponed and what already applies
• Metarix: Grok tool page
• Metarix: SpaceXSI company hub