# How Claude Handles Information About xAI Products, SuperGrok, and API Subscriptions

> Discover how Claude handles xAI product information, including SuperGrok and API subscriptions, through its guardrail layer and approved whitelist. Learn about its restrictions on pricing and unreleased features.

- Repository: [Ásgeir Thor Johnson/system_prompts_leaks](https://github.com/asgeirtj/system_prompts_leaks)
- Tags: deep-dive
- Published: 2026-02-16

---

**Claude uses a dedicated guardrail layer defined in [`xAI/grok-4.md`](https://github.com/asgeirtj/system_prompts_leaks/blob/main/xAI/grok-4.md) to restrict discussions about xAI products to an approved whitelist, redirect pricing queries to official pages, and prevent hallucinations about unreleased features.**

The leaked system prompts from the `asgeirtj/system_prompts_leaks` repository reveal how Anthropic engineers have hardcoded specific rules governing Claude's behavior when users ask about xAI's Grok models, SuperGrok subscriptions, or API access. These constraints operate as a policy layer that filters the model's generative output before it reaches the user.

## The xAI Knowledge Block in Claude's System Prompts

Located at [`xAI/grok-4.md`](https://github.com/asgeirtj/system_prompts_leaks/blob/main/xAI/grok-4.md) in the leaked repository, the xAI knowledge block functions as a controlled vocabulary for Claude's responses. Unlike general knowledge retrieved during training, this block is injected directly into the context window, giving it precedence over the model's parametric memory.

### Product Scope Limitations

Claude is explicitly restricted to mentioning only approved xAI products listed in the system prompt. According to lines 10-13 of [`grok-4.md`](https://github.com/asgeirtj/system_prompts_leaks/blob/main/grok-4.md), the allowed set includes **Grok 4**, **Grok 3**, the **Grok iOS/Android apps**, and the **X mobile apps**. Any product outside this whitelist—such as the rumored "BigBrain mode" referenced in [`xAI/grok-3.md`](https://github.com/asgeirtj/system_prompts_leaks/blob/main/xAI/grok-3.md)—is strictly prohibited. If a user inquires about unlisted features, Claude must decline to answer rather than speculate.

### Subscription-Aware Response Rules

The system prompt contains specific instructions regarding **SuperGrok**, xAI's premium subscription tier. As defined in lines 13-15 of [`grok-4.md`](https://github.com/asgeirtj/system_prompts_leaks/blob/main/grok-4.md), Claude knows that SuperGrok is the sole paid plan that upgrades a user's Grok 3 quotas. However, the model is explicitly forbidden from fabricating specific pricing figures or rate limits. Instead, per lines 15-16, Claude must redirect users to the official pricing page at `https://x.ai/grok` to obtain current cost information.

### API Documentation Guidance

When handling inquiries about the **xAI API**, Claude follows a strict referral protocol. Line 18 of [`grok-4.md`](https://github.com/asgeirtj/system_prompts_leaks/blob/main/grok-4.md) instructs the model to direct users to the API portal URL without elaborating on technical specifications, authentication methods, or endpoint details. This ensures that users receive the most current documentation directly from the source rather than potentially outdated information embedded in the model's training data.

### Knowledge Freshness and Presentation Standards

The xAI knowledge block explicitly declares that Claude's information is **"continuously updated"** and operates with **no fixed knowledge cutoff** for xAI products (line 21). This instruction elevates the system prompt's authority over the model's base training knowledge. Additionally, lines 22-23 mandate that Claude use **tables** when presenting product comparisons, enforcing a consistent UI pattern across all xAI-related responses.

## How the Guardrail Layer Prevents Hallucinations

The architecture described in [`grok-4.md`](https://github.com/asgeirtj/system_prompts_leaks/blob/main/grok-4.md) functions as a **pre-generation filter**. When a user query triggers the xAI topic classifier, Claude does not immediately generate a response based on its parametric knowledge. Instead, it retrieves the xAI knowledge block from the system prompt and validates the query against the allowed product whitelist and response rules.

If the query concerns SuperGrok pricing, the guardrail blocks any attempt to generate specific dollar amounts and instead inserts the standardized redirect to `x.ai/grok`. If the query mentions unapproved products like "BigBrain mode" (explicitly marked as unavailable in [`xAI/grok-3.md`](https://github.com/asgeirtj/system_prompts_leaks/blob/main/xAI/grok-3.md)), the guardrail triggers a refusal response rather than allowing the model to hallucinate capabilities. This mechanism ensures that Claude's xAI product information remains accurate, current, and legally compliant.

## Code Example: Querying Claude About xAI Products

The following Python example demonstrates how the xAI guardrails automatically activate when querying Claude through the Anthropic API. Note that the application code does not need to implement separate validation logic—the guardrails embedded in the system prompt handle compliance automatically.

```python
import anthropic
import os

# Initialize client with API key from environment

client = anthropic.Anthropic(api_key=os.getenv("ANTHROPIC_API_KEY"))

def query_claude_about_xai(user_question: str) -> str:
    """
    Submit a question about xAI products to Claude.
    The xAI guardrail in the system prompt automatically
    enforces response constraints.
    """
    response = client.messages.create(
        model="claude-sonnet-4-20250514",
        max_tokens=1024,
        temperature=0.0,
        messages=[{"role": "user", "content": user_question}]
    )
    return response.content[0].text

# Example 1: Pricing query (triggers redirect guardrail)

pricing_question = "How much does SuperGrok cost per month?"
print(f"Q: {pricing_question}")
print(f"A: {query_claude_about_xai(pricing_question)}")

# Expected: Redirect to https://x.ai/grok instead of stating price

# Example 2: API inquiry (triggers documentation redirect)

api_question = "How do I authenticate with the xAI API?"
print(f"\nQ: {api_question}")
print(f"A: {query_claude_about_xai(api_question)}")

# Expected: URL to API portal without technical implementation details

```

## Related System Prompt Files

The `asgeirtj/system_prompts_leaks` repository contains several files governing Claude's xAI-related behavior:

- **[`xAI/grok-4.md`](https://github.com/asgeirtj/system_prompts_leaks/blob/main/xAI/grok-4.md)**: Primary guardrail for Grok 4, SuperGrok subscriptions, and current API information.
- **[`xAI/grok-4.1-beta.md`](https://github.com/asgeirtj/system_prompts_leaks/blob/main/xAI/grok-4.1-beta.md)**: Beta version guardrail maintaining identical SuperGrok and API handling rules.
- **[`xAI/grok-3.md`](https://github.com/asgeirtj/system_prompts_leaks/blob/main/xAI/grok-3.md)**: Legacy guardrail for Grok 3, explicitly noting that "BigBrain mode" is unavailable to any plan.
- **[`Anthropic/claude-opus-4.6.md`](https://github.com/asgeirtj/system_prompts_leaks/blob/main/Anthropic/claude-opus-4.6.md)**: General product description file useful for understanding how Claude describes its own capabilities in contrast to xAI products.

## Summary

- Claude relies on a dedicated **xAI knowledge block** in [`xAI/grok-4.md`](https://github.com/asgeirtj/system_prompts_leaks/blob/main/xAI/grok-4.md) to govern all responses about xAI products, overriding general training knowledge.
- The model maintains a strict **product whitelist** that includes Grok 4, Grok 3, and mobile apps, while explicitly prohibiting mention of unverified features like "BigBrain mode."
- **SuperGrok** is the only recognized paid tier, but Claude **cannot state pricing** and must redirect users to `https://x.ai/grok`.
- All **API inquiries** trigger a standardized redirect to the official API portal without technical elaboration.
- The guardrail architecture operates as a **pre-generation filter**, preventing hallucinations about pricing, capabilities, or unreleased products.

## Frequently Asked Questions

### What is SuperGrok and how does Claude describe it?

SuperGrok is xAI's premium subscription tier that upgrades user quotas for Grok 3 access. According to the system prompt in [`xAI/grok-4.md`](https://github.com/asgeirtj/system_prompts_leaks/blob/main/xAI/grok-4.md), Claude recognizes SuperGrok as the only paid plan that modifies Grok 3 usage limits, but the model is explicitly forbidden from detailing specific quota numbers or pricing tiers.

### Why does Claude redirect users to x.ai for pricing instead of stating prices directly?

The guardrail defined in lines 15-16 of [`xAI/grok-4.md`](https://github.com/asgeirtj/system_prompts_leaks/blob/main/xAI/grok-4.md) explicitly prohibits Claude from fabricating pricing figures or subscription limits. This prevents the model from hallucinating outdated or incorrect cost information that could mislead users. Instead, Claude must provide the canonical URL `https://x.ai/grok` to ensure users receive real-time, accurate pricing data directly from xAI.

### How does Claude handle questions about the xAI API?

When users inquire about the xAI API, Claude follows the protocol established in line 18 of [`xAI/grok-4.md`](https://github.com/asgeirtj/system_prompts_leaks/blob/main/xAI/grok-4.md). The model redirects the user to the official API portal URL without providing implementation details, authentication schemas, or endpoint specifications. This ensures that developers receive the most current technical documentation rather than potentially deprecated information from the model's training data.

### What prevents Claude from inventing details about unreleased xAI features?

The system prompt maintains a strict **product whitelist** (lines 10-13 of [`grok-4.md`](https://github.com/asgeirtj/system_prompts_leaks/blob/main/grok-4.md)) that explicitly limits Claude to discussing only verified xAI products: Grok 4, Grok 3, and the official mobile applications. Features not on this list—such as the rumored "BigBrain mode" explicitly marked as unavailable in [`xAI/grok-3.md`](https://github.com/asgeirtj/system_prompts_leaks/blob/main/xAI/grok-3.md)—trigger a refusal response. This whitelist acts as a hard constraint that prevents the generative model from hallucinating capabilities or products that do not exist.