Unveiling Grok 4.7

SpaceXAI | News Update | September 21, 2026

"Grok 4.7 represents our most sophisticated leap forward in the realms of professional knowledge work and complex software engineering."

Grok 4.7 has arrived, serving as the new gold standard for users requiring high-level coding capabilities and deep intellectual labor. This model is engineered to persevere through grueling tasks, implement more rigorous self-correction, and operate under the most precisely calibrated safety frameworks we have ever deployed.

Despite these upgrades, it maintains the same velocity and cost structure as Grok 4.6, ensuring it remains a top-tier contender in terms of efficiency.

On the CursorBench 4.0 metric—which specifically tests the model's endurance during protracted coding sessions—Grok 4.7 sits at the absolute frontier of price-to-performance.

---

🛠️ Architectural Enhancements

The transition from the previous iteration to Grok 4.7 involved several critical upgrades:

• Expanded Foundation: It utilizes a larger base model than its predecessor.
• Rigorous Training: The model underwent an extended reinforcement learning (RL) phase focusing on a "hard" dataset, specifically prioritizing problems that require hours of computation to solve.
• Cognitive Management: Enhanced ability to verify its own outputs and handle expansive context windows.
• Native Integration: It was trained to inherently understand the Grok Bot harness, optimizing it for general knowledge tasks and fluid conversation.

---

📊 Performance Metrics

Grok 4.7 demonstrates significant gains across various professional domains. Below is the detailed breakdown of how it compares to other frontier models:

*Denotes high-effort mode.

Professional Knowledge Analysis

In evaluations like GDPval and AA Briefcase—which simulate the workflows of financial analysts, nurses, and attorneys—Grok 4.7 shows a marked improvement over Grok 4.6 and remains competitive with the industry's leading models.

The Elo rating for professional knowledge can be expressed as: $\text{Elo}_{\text{Grok 4.7 (xhigh)}} = 1695$ (Compared to Fable 5.1 Max at $1735$ and GPT-6 Astra Max at $1542$).

---

🛡️ Safety & Cybersecurity

Grok 4.7 introduces a completely redesigned safeguard architecture. It is currently our most resilient model regarding jailbreak attempts and refusal accuracy.

Key Safety Highlights:

1. Dual-Use Balance: In fields like biology and cyber-defense, it maximizes utility for helpful tasks while strictly blocking harmful ones.
2. Biosafety: It achieved a score of $62.4\%$ on the LatchBio benchmark.
3. Cyber Defense: On HackerBench v0.3, it allows only $3.3\%$ of malicious prompts to pass, while maintaining a low refusal rate for legitimate security research.

Note: We have initiated a limited, invite-only program for cybersecurity partners to utilize Grok 4.7's red-teaming capabilities for defensive research.

---

💰 Pricing & Access

Grok 4.7 is available immediately via the following channels:

• Integrations: Cursor and Grok Build.
• Infrastructure: Grok API, cloud platforms, model routers, and 3rd-party coding harnesses.

Cost Structure:

• Standard: \text{Input} = \2 / 10^6 \text{ tokens} \mid \text{Output} = $6 / 10^6 \text{ tokens}$
• Fast Variant: $2\times$ Speed $\rightarrow$ $2\times$ Price.

🚀 Get Started

To begin integrating Grok 4.7 into your workflow, follow these steps:

• Visit x.ai/build
• Generate an API key in the Console
• Review the technical documentation at docs.x.ai
• Test the model for free in Grok Build

// Example API Configuration
{
  "model": "grok-4.7",
  "temperature": 0.7,
  "max_tokens": 4096,
  "top_p": 1.0
}