Grok API watermarks seven models by default across all endpoints

Seven Grok API models now carry an invisible statistical watermark by default across every SpaceXSI endpoint, a change effective Nov. 7. SpaceXSI’s developer documentation says the default exists “to comply with the EU AI Act.” The mark covers ordinary responses, structured outputs and tool-call arguments; only team administrators can reverse it.

Which Grok models and endpoints are covered?

The rollout covers four numbered models:

• grok-4.7
• grok-4.6
• grok-4.5
• grok-4.3

Three additional covered models are the grok-4.20 reasoning, non-reasoning and multi-agent variants. The list fixes the model scope of the default, while SpaceXSI applies the setting across the service rather than limiting it to one product surface.

SpaceXSI says every API endpoint receives the setting, including the global api.x.ai endpoint and any regional endpoints. It applies to both the Responses API and the Chat Completions API. Structured outputs and tool-call arguments receive the mark along with ordinary text responses.

Because both API styles and both output classes sit in scope, an application can receive marked data without sending a special flag. The absence of request-level controls means the watermark follows account and team configuration, not a per-call instruction from client code.

How does the mark work without adding characters?

No extra characters are added to the output. Instead, SpaceXSI says the watermark “slightly shapes which tokens the model picks,” leaving no extra characters, markup or hidden text in the response. The exact token sequence can therefore differ from an unmarked generation, but the client still receives a normal model response.

Detection follows the same statistical approach. SpaceXSI scores a passage and reports one of three states:

• watermarked
• not watermarked
• uncertain

Longer text gives the detector a stronger signal. The uncertain state preserves a third outcome when the statistical score does not support a binary finding.

Access to the detection API is granted by application. SpaceXSI also says the watermark does not change response quality or speed, leaving the added provenance signal as the stated functional difference for API customers.

What changes for API integrations?

Existing integrations do not need code changes, SpaceXSI says. There is no request parameter, and the response format remains the same apart from a new x-watermark header. A value of true means that request’s output was watermarked. Any other value, or no header at all, means it was not.

The header gives developers a per-request status without changing the body of the response. It also exposes the practical result of having no request parameter: clients cannot enable or disable the mark for an individual call through the API itself.

SpaceXSI’s documentation includes JavaScript, Python and curl examples for reading the header. The examples include the OpenAI SDKs pointed at SpaceXSI’s base URL, so developers can use familiar client code while inspecting the new response field.

Who can turn watermarking off?

The exception policy sits at team level rather than request level. Only team admins see the “Model output watermarking” row in the SpaceXSI Console, where they can switch the mark off model by model. The setting applies to every API key on the team and reaches new requests within five minutes.

When admins opt out, the Console asks them 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.” The confirmation assigns responsibility for the resulting marking and disclosure obligations to the administrator.

Some models may be locked on and cannot be switched off from the Console. SpaceXSI records each opt-in and opt-out in the team’s audit log. The structure separates provider control from team control: SpaceXSI can require watermarking on selected models, while an admin can disable it on others and leave a recorded account of the decision.

Because one setting covers every API key, a team does not need separate policies for each credential. The Console’s five-minute propagation window gives the change time to reach new requests, rather than requiring clients to alter their request syntax.

How do OpenAI and Anthropic compare?

The three providers now expose different controls over text provenance. SpaceXSI has made watermarking the default for seven API models. OpenAI has left its API feature off by default while planning a regional rollout for other products. Anthropic applies watermarking worldwide across supported Claude API and app output.

On Oct. 5, OpenAI said API customers globally could opt in to text watermarking for select models. It said the feature “will remain off by default in the API.” That differs from SpaceXSI’s API default and leaves the customer responsible for opting in.

OpenAI’s plan called for ChatGPT and Codex text in the European Union to receive the mark over the following weeks, WebProNews reported at the time. The plan applied to product surfaces outside the API and set a narrower regional scope than Anthropic’s worldwide policy.

Anthropic also applies watermarking worldwide. In August, the company said future Claude models would carry a watermark based on Google DeepMind’s SynthID-Text method. Anthropic described the method as adding no hidden characters. Its help center says marks apply to output from supported models across the Claude API and apps “wherever Claude is offered, worldwide.”

That worldwide coverage puts Anthropic closer to SpaceXSI’s broad rollout than OpenAI’s API policy. The three companies still differ on activation: SpaceXSI enables the mark automatically for covered models, OpenAI requires an API opt-in, and Anthropic has committed supported Claude output to worldwide marking.

Who can detect watermarked text?

Access to text detection remains gated at all three model providers:

• OpenAI limits its text detector to approved researchers and expert organizations.
• Anthropic’s detector remains in private preview for eligible organizations.
• SpaceXSI grants access to its detection API by application, with requests handled through a form.

Google opened its public SynthID Detector in October. The tool checks images, video and audio files, giving developers a public verification option for those media types. Its public access does not extend to the text detectors operated by OpenAI, Anthropic and SpaceXSI.

The separation between marking and verification has practical consequences. A provider can apply a watermark to every eligible response while reserving detection for researchers, approved organizations or vetted applicants. Coverage of generated text therefore does not guarantee that an API customer can independently test it.

For teams deploying the seven affected Grok models, the next operational choice is administrative: retain the EU-compliance default or document an applicable legal exception and switch the mark off, with each change recorded in the team’s audit log.