Grok API Watermark Goes Live Today: 7 Models Marked by Default

From today, November 7, 2026, every text answer that comes back from seven of xAI's Grok models carries an invisible watermark. You do not have to change your code, your prompts, or your API keys. The mark is on by default on every endpoint, and the only visible sign of it is a new response header. If you build on the Grok API, this is a change that reaches your product whether you asked for it or not, so it is worth knowing exactly which models it touches, what it costs, and how to turn it off if the law allows you to.

The change comes from xAI's developer documentation, which now carries the name SpaceXSI, the SpaceX unit that builds Grok. The docs page, last updated October 31, is the whole announcement: no blog post, no release-notes entry. The reason given is one line: SpaceXSI watermarks text model output to comply with the EU AI Act, whose deadline for marking AI-generated text falls on December 2, 2026, twenty-five days from today.

What changed today

Starting November 7, 2026, responses from seven Grok models include a watermark, on by default, with an opt-out per model. The seven are 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. Watermarking applies to text output from text generation requests, on every text generation API, including the Responses API and the Chat Completions API. It covers the generated text in the response: ordinary text, structured outputs, and tool call arguments.

The mark is enabled on every SpaceXSI API endpoint, including the global endpoint at api.x.ai and any regional endpoint. This is not a European setting or a per-key toggle: a team in the United States gets marked text, and so does a team calling a regional endpoint elsewhere. The default applies to the whole team, and the opt-out is a team-level decision.

Nothing about the request changes. There is no parameter to send, and the response format is unchanged apart from one addition: the x-watermark response header. A value of true means the output for that request was generated with the watermark. Any other value, or no header at all, means it was not.

The seven models, and what each one costs

The watermark itself costs nothing. xAI's docs say it does not change response quality or speed, and there is no line item for it on the pricing page. What you pay is the ordinary per-token price of the model you call. Those prices, read live from xAI's pricing page on November 7, 2026, are:

Three pieces of fine print apply to every row. First, the rates above are for the standard US regions; one US region on the same pricing page lists grok-4.7 slightly higher, at $2.20 input and $6.60 output per million tokens. Second, long context costs more: once a request passes 200,000 tokens, the input and output rates double, so grok-4.7 becomes $4.00 in and $12.00 out. Third, the batch API carries a 20% discount.

To put the numbers in practical terms, a workload of 10 million input tokens and 2 million output tokens costs $32.00 on grok-4.7 (about £23.91 or C$45.01), and $17.50 on grok-4.3 or any of the grok-4.20 models (about £13.07 or C$24.62). Cached input, where the model reuses prompt tokens it has already seen, is where the bill shrinks: grok-4.7 charges $0.50 per million cached input tokens instead of $2.00. Currency conversions use November 7, 2026 mid-market rates of 1 USD = £0.7471 and C$1.4066, and are approximations; xAI bills in US dollars.

For context, per-token prices have fallen sharply this year. OpenAI and Anthropic cut their own prices by up to 80% over the last four months, so these Grok rates are mid-pack rather than a bargain. The watermark adds no charge on top of any of this, which is the one piece of pricing news that is entirely good.

How the watermark actually works

The watermark is not visible. It is applied while the model generates text: it slightly shapes which tokens the model picks, so it adds no characters, markup, or hidden text to the response. In practice, the text you receive reads exactly as it would have read without the mark: the signal lives in the pattern of word choices, not in anything a reader can see.

Detection is statistical. SpaceXSI scores a piece of text against the watermark and reports whether it is watermarked, not watermarked, or uncertain. Longer text gives a stronger signal, which means short replies may not carry enough evidence for a confident call either way. Access to watermark detection is granted by application; there is no public checker, and the docs do not publish detection accuracy figures of their own.

Two limits follow from that design, and both are worth stating plainly. First, the docs say nothing about how well the mark survives editing or paraphrase, so a passage that has been heavily rewritten may no longer test positive. OpenAI, by contrast, published its numbers: at a 1% false positive rate, its textGrain detector found the mark in about 80% of 200-token passages and about 95% of 400-token passages. xAI has published no equivalent figures, so treat its mark as a signal, not a guarantee. Second, a missing mark proves nothing on its own. Text generated before November 7 carries no mark, text from a model you opted out carries none, and text from another provider's model carries a different mark or none at all.

How to check a response, and how to opt out

Checking is the easy part. Read the x-watermark header on the Responses API or Chat Completions API response. If it says true, that request's output was generated with the watermark. The docs include a short code sample that logs the header alongside the generated text, and no request parameter is needed to make the header appear.

Opting out takes more doing, and it is deliberately a team decision. Only team admins can change the setting, and only in the SpaceXSI Console. The steps, as the docs describe them, are:

1. Open the Team settings page and find the Model output watermarking row. It shows On (all models), Off (all models), or Custom. Only team admins see this row.
2. Select the settings icon next to it, then clear the checkbox for each model you want to opt out. Select all turns every model on or off at once.
3. Select Save. If you turned a model off, confirm with Opt Out.

Changes apply to new requests within five minutes. After you opt a model out, its new requests stop including the watermark; after you opt back in, they start including it again. The setting applies to every API key on the team, and every opt-in and opt-out is recorded in the team's audit log, which team admins can view in the Console.

Two conditions come with the opt-out. First, opting out asks you to confirm that an exception under applicable law applies to your use, and that you are solely responsible for any marking, disclosure, or other legal requirement. In other words, switching the mark off moves the compliance duty from xAI to you. Second, some models may be locked on for your team. Locked models are listed in the dialog but cannot be turned off from the Console, and the docs direct you to your team admin or to xAI support if you need one switched off.

What the watermark does not cover

The docs cover API text only, and the gaps are as important as the coverage. The watermark does not apply to the Grok apps, to grok.com, or to Grok on X. It does not apply to images or video made with Grok Imagine. It does not apply to text generated before November 7, which carries no mark even if you test it now. And it does not apply to models outside the seven listed: grok-4.20's three variants are in, but no other model is named.

One question stays open: cloud partners. The docs list Google Cloud Vertex AI and Microsoft Foundry as platforms that offer Grok, but their pages do not mention the watermark, and the docs do not say whether calls through those platforms are marked. If you reach Grok through a cloud partner, check with that provider before assuming your output is marked.

Finally, the watermark says nothing about who wrote a passage or whether it is true. A watermarked passage can still be wrong, misleading, or heavily edited by a person, and an unmarked passage can still be machine-generated. OpenAI makes the same point about its own mark: a watermark does not measure human contribution, establish ownership, identify the user, or verify accuracy, and the absence of a detected watermark does not prove human authorship.

How xAI compares with OpenAI and Anthropic

xAI is not the first provider to switch text watermarking on. OpenAI announced its approach on October 5, and Anthropic expanded its own marking on November 3. What differs is the default each one chose.

OpenAI announced its approach on October 5, 2026, in a post titled "Our approach to EU text provenance rules." Its method, called textGrain, adds an invisible statistical signal to the model's word choices, and OpenAI plans to release it as open source. The API side is opt-in: customers turn it on under Organization settings, and it stays off unless they do. For its own products, OpenAI is adding the mark to eligible ChatGPT and Codex text in the European Union over the coming weeks, on all plans, and says it is not making text watermarking a global default at launch. OpenAI had a busy November on the API side in general: its GPT-6 Cyber model, built for vulnerability research, shipped the same week.

Anthropic moved on November 3, 2026, when its Claude Platform release notes added the watermark to Claude Mythos 5, Claude Opus 4.7, and Claude Sonnet 4.6. An email to customers had named a fourth model, Claude Opus 4.6, which does not appear in the release note. Newer models, including Claude Opus 5.5, already carried the mark from launch. Anthropic's method is a version of SynthID-Text, the technique Google DeepMind published in Nature in 2024, and it applies worldwide because Anthropic says it has no durable way to limit the mark by region. One quirk applies only to Sonnet 4.6: a request that sets temperature, top_p, or top_k to a non-default value might not carry the mark, and Anthropic's docs advise leaving all three out if you rely on it. Bedrock rollout for the three models completes by November 10.

The contrast between the three is the story. OpenAI leaves the choice with the customer and keeps the API default off. Anthropic gives the customer no choice at all. xAI sits between them: on by default, but removable by a team admin who is willing to sign for the legal responsibility. For a developer, that middle position is the one that demands an actual decision this week, rather than a setting you can ignore.

The law behind the switch

All three labs point at the same law. Article 50(2) of the EU AI Act requires providers of AI systems that generate synthetic text, audio, image, or video content to ensure the outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. The marking rules have applied since August 2, 2026. Systems already on the market before that date get a transition period: under Article 111(4), as amended by the Digital Omnibus regulation, they must comply by December 2, 2026.

That date explains the timing. xAI's November 7 switch lands twenty-five days before the deadline, OpenAI's October 5 announcement came about two months before it, and Anthropic's November 3 expansion came a month before it. None of the three waited for the last day, and none of them treats the mark as optional. The differences are in scope and default, not in whether to mark at all.

One caution: the AI Act's marking rule has exceptions, and the docs do not assess them. xAI's opt-out asks you to confirm that an exception under applicable law applies to your use. If you are not sure one does, that confirmation is not a formality to click through: it is a legal statement, recorded in your team's audit log.

What developers should do this week

Most teams need an hour, not a project, to deal with this change. Start by checking which of the seven models your product actually calls: if you call grok-4.7, grok-4.6, grok-4.5, grok-4.3, or any grok-4.20 variant through the API, your output is marked from today, and if you call an older or unlisted model, it is not.

Next, read the x-watermark header on a few live responses and log what you see. It costs nothing and confirms the docs match your traffic. If your product stores model output, keep the header value with the stored text: the mark itself needs a detector to read, while the header is a plain fact about that request.

Then decide about the opt-out, and make it a deliberate team decision rather than one person's click. If your use falls under a legal exception, document the reasoning, have a team admin switch the models off, and note the change in your own records alongside xAI's audit log. If it does not, leave the mark on and tell your users what it means. And if you reach Grok through Google Cloud or Microsoft Foundry, ask that provider whether your output is marked, because the docs do not answer that question.

Finally, update anything your users see. If your product tells users that its text is AI-generated, the watermark supports that claim. If it does not, the mark does not fix the gap on its own: a statistical signal that only xAI's detector can read is not a disclosure to your users.

Three things not to assume

First, do not assume the mark survives editing. xAI says nothing about how well the mark holds up to paraphrase, and OpenAI's published numbers show how fast a mark fades: replacing a quarter of the words with synonyms cut its detection rate from about 92% to 17%. A passage that has been through a human editor, a translation, or a rewrite may test clean even though a model wrote the first draft.

Second, do not assume a missing mark means a human wrote the text. Short passages carry too little signal to test, pre-November 7 text was never marked, opted-out models produce unmarked text, and other providers' models carry different marks or none. A negative result is the absence of evidence, not evidence of absence.

Third, do not assume the watermark settles your own compliance duties. The mark is xAI's answer to its obligations as a model provider. Your obligations as the party putting the text in front of users are separate, and may include disclosure, record-keeping, or marking of your own. Anthropic's help center makes the same point: the mark on Claude's output does not settle your Article 50 duties for you.

Is my Grok API output watermarked now?

If you call grok-4.7, grok-4.6, grok-4.5, grok-4.3, or any grok-4.20 variant through the API, yes, from November 7, 2026, unless a team admin opted that model out. Check the x-watermark response header: true means that request's output was watermarked.

Does the watermark cost extra?

No. There is no charge for watermarking, and it does not change the per-token price of any model. You pay the ordinary API rates, starting at $1.25 per million input tokens for grok-4.3 and the grok-4.20 models, and $2.00 for grok-4.7, grok-4.6, and grok-4.5.

Can I turn the watermark off?

Yes, if you are a team admin. In the SpaceXSI Console, under Team settings, the Model output watermarking row lets you switch models off per model. You must confirm that a legal exception applies to your use and that you are solely responsible for marking and disclosure. Some models may be locked on for your team.

How do I check whether a response was watermarked?

Read the x-watermark header on the API response. A value of true means the output was generated with the watermark. Full detection, which scores a passage and reports watermarked, not watermarked, or uncertain, is available by application only.

Does the watermark affect quality or speed?

According to xAI's docs, no. The mark slightly shapes which tokens the model picks and adds no characters, markup, or hidden text. OpenAI published benchmark comparisons for its own watermark showing no meaningful performance difference, and xAI makes the same claim, though it has published no benchmark figures of its own.

Does the watermark cover the Grok app or posts on X?

No. The docs cover API text only. They say nothing about the Grok apps, grok.com, Grok on X, or images and video from Grok Imagine. Those surfaces are unmarked as far as the published docs show.

The watermark is a small change with a large default. Nothing in your code breaks, nothing in your bill changes, and nothing your users see is different. What changes is that every Grok API reply now carries a machine-readable claim about where it came from, and turning that claim off is a decision with a name attached to it in an audit log. For most teams, the right move is the boring one: confirm which models you call, log the header, leave the mark on, and tell your users what it means. The deadline that drove all of this is December 2, and xAI got there twenty-five days early.