Grok API will add invisible watermarks to text from seven models by default from November 7

SpaceXSI, the company formerly known as xAI, will start watermarking text generated through the Grok API by default on November 7, 2026. The change is described on a new "Output Watermarking" page in the official developer documentation, last updated on October 31, 2026, and it covers seven models: grok-4.7, grok-4.6, grok-4.5, grok-4.3 and the three grok-4.20 variants (reasoning, non-reasoning and multi-agent). From that date, text returned by these models will carry a statistical mark that the company can later use to identify the output as machine-generated.

The documentation states that the watermark is applied to comply with the EU AI Act and that it is enabled on every SpaceXSI API endpoint, including the global endpoint at api.x.ai and the regional endpoints. The scope is wide. It covers the Responses API and the Chat Completions API, and within those it applies to ordinary text, to structured outputs and to the arguments the model writes when it calls a tool. Developers do not need to change their requests, because there is no parameter to send and the response body keeps its format.

The only visible sign of the change is a new response header, `x-watermark`, which returns `true` when the output was watermarked. Team administrators can switch the feature off model by model in the SpaceXSI Console, but only after confirming that a legal exception applies to them and accepting sole responsibility for any marking or disclosure obligations that remain.

··········

HOW GROK API OUTPUT WATERMARKING WORKS.

According to the documentation, the watermark is applied during generation. It slightly influences which tokens the model selects, so it adds no characters, no markup and no hidden text to the response. A string copied from a watermarked Grok answer therefore looks exactly like any other string; there are no zero-width characters to strip and nothing that a text sanitizer would detect.

Detection works statistically. SpaceXSI says its detector returns one of three verdicts, "watermarked", "not watermarked" or "uncertain", and that longer passages give a stronger signal. That last point follows from the method: a token-selection bias becomes measurable only across many tokens, so a one-line reply or a short JSON object will often fall into the uncertain band. The company has not published accuracy figures, false-positive rates or results after paraphrasing or translation, which are the usual weaknesses of statistical text watermarks.

Access to the detection API is not open. The page says it is granted by application through a form, and it does not say who qualifies, whether detection will be offered to the public or whether it will be priced. For now, the people who can check whether a given text came from a watermarked Grok model are SpaceXSI and the organizations it admits to that program.

SpaceXSI also states that watermarking does not affect response quality or speed. That is a vendor claim. The company has not released evaluations comparing watermarked and unwatermarked outputs on benchmarks, structured-output validity or tool-calling accuracy.

··········

WHICH GROK MODELS AND ENDPOINTS ARE COVERED.

The seven models named in the documentation span the current Grok API catalog. The table below lists them with the context window and standard prices published on the official models page, as checked on November 1, 2026. Prices are unchanged by the watermarking policy and are shown here to identify the models.

........

Model (API name)	Context	Input, USD per 1M tokens	Output, USD per 1M tokens	Watermark by default from Nov 7
grok-4.7	500K	$2.00	$6.00	Yes
grok-4.6	500K	$2.00	$6.00	Yes
grok-4.5	500K	$2.00	$6.00	Yes
grok-4.3	1M	$1.25	$2.50	Yes
grok-4.20-0309-reasoning	1M	$1.25	$2.50	Yes
grok-4.20-0309-non-reasoning	1M	$1.25	$2.50	Yes
grok-4.20-multi-agent-0309	1M	$1.25	$2.50	Yes
grok-build-0.1	256K	$1.00	$2.00	Not listed

........

Input and output prices are those for prompts under 200,000 tokens; above that threshold the official page doubles both. The watermarking page refers to the grok-4.20 models without the "-0309" date suffix used on the models page. The coding model grok-build-0.1 appears on the models page but is absent from the watermarking list, and the documentation does not explain whether it is exempt or simply not yet included.

A broader comparison of Grok 4.7 prices with rival flagship models is available in our analysis of GPT-6.1 Terra, Claude Sonnet 5.5, Gemini 4 Argon and Grok 4.7 API pricing.

The policy applies to text generation only. The documentation does not extend it to Grok's image, video or voice products, which are governed by separate pages.

··········

WHAT CHANGES FOR DEVELOPERS USING THE GROK API.

For most applications nothing breaks on November 7. Requests stay the same, the response schema stays the same and the content of answers does not gain any extra characters. The practical questions concern disclosure, logging and agent pipelines.

The first decision is whether to keep the default. Teams that publish Grok-generated text to end users in the EU, or that sell products which do, will generally want the watermark on, since machine-readable marking is exactly what the AI Act asks generative AI providers to supply. Teams that use Grok for internal processing, classification or code generation may prefer to opt out, but the console flow makes them confirm that a legal exception applies before it accepts the change.

The opt-out procedure, as documented, runs as follows:

........

Step	What happens
Open Team settings in the SpaceXSI Console	The "Model output watermarking" row shows On (all models), Off (all models) or Custom
Open the row's settings	Each model has its own checkbox, plus a "Select all" toggle
Clear the models to opt out and save	Turning a model off requires confirming with "Opt Out"
Wait for propagation	New requests reflect the change within 5 minutes
Audit	Every opt-in and opt-out is recorded in the team audit log

........

The setting is team-wide and applies to every API key in the team, so a company cannot keep watermarking for a customer-facing product while disabling it for an internal tool under the same team. Running those workloads under separate teams is the practical way to get different behavior. The page also warns that some models may be locked on for a given team; those appear in the dialog without a checkbox, and the documentation directs affected users to their team admin or to support.

Applications that log responses can record the `x-watermark` header alongside each completion, which gives an audit trail of which outputs were marked. The documentation includes examples for the AI SDK, the OpenAI SDK in JavaScript and Python, and curl. The snippet below is an illustrative sketch using the OpenAI Python SDK's raw-response interface against the Grok API base URL.

from openai import OpenAI

client = OpenAI(api_key="XAI_API_KEY", base_url="https://api.x.ai/v1")
raw = client.chat.completions.with_raw_response.create(
    model="grok-4.7",
    messages=[{"role": "user", "content": "Summarize the AI Act transparency rules."}],
)
completion = raw.parse()
watermarked = raw.headers.get("x-watermark") == "true"
print(watermarked, completion.choices[0].message.content)

Agent builders have the most to test. Because the watermark also applies to structured outputs and to tool-call arguments, it touches the parts of a response that software parses rather than people read. SpaceXSI says quality is unaffected, but teams running JSON-heavy workflows or long chains of function calls should compare schema-validation failures and tool-selection errors before and after November 7, at least until independent measurements exist.

··········

THE EU AI ACT RULE BEHIND GROK API WATERMARKING.

The legal basis cited by SpaceXSI is the transparency chapter of the EU AI Act. According to the European Commission's published summary of the rules, which apply from 2 August 2026, providers of generative AI systems must apply a machine-readable mark to synthetic content generated or manipulated by AI and must make that mark detectable. The obligation covers text as well as images, video and audio.

The Commission lists exceptions. Marking is not required when an AI system performs an assistive function for standard editing, or when it does not substantially alter the input data supplied by the deployer or its meaning. Those exceptions explain the shape of the opt-out in the SpaceXSI Console, where an administrator must state that one of them, or another legal exception, applies.

The Commission's page also describes a grace period. For generative AI systems placed on the market before 2 August 2026, the marking obligation under Article 50(2), as amended by the AI Omnibus, applies only from December 2026. A default switch on November 7 lands about a month before that grace period ends for older systems. Fines for breaching the transparency obligations can reach €15 million or 3% of worldwide annual turnover, according to the same Commission summary.

Compliance can also be demonstrated through the voluntary Code of Practice on Transparency of AI-generated Content, which the Commission published on 10 June 2026. SpaceXSI's documentation does not say whether the company has signed it, and it does not say whether the watermark follows any shared technical standard that third-party detectors could read.

··········

GROK API WATERMARKING: CONFIRMED FACTS AND OPEN QUESTIONS.

The documented facts are specific. From November 7, 2026, seven Grok text models will watermark their output by default on all API endpoints, across the Responses and Chat Completions APIs, including structured outputs and tool-call arguments. The mark is statistical, invisible and signalled by an `x-watermark: true` header, and team administrators can opt out per model only by asserting a legal exception, with every change logged.

Several points remain undisclosed. SpaceXSI has published no detection accuracy, no results on how the mark survives paraphrasing or translation, and no criteria for access to the detection API. The page does not explain why grok-build-0.1 is absent from the list or which models may be locked on for particular teams. Its statement that quality and speed are unaffected has not been checked independently, which leaves agent and structured-output workloads as the area where developers will need their own measurements once the default takes effect.

··········

FOLLOW US FOR MORE.

·····

DATA STUDIOS

·····

datastudios.org