Introducing Grok 4.7
Key Points
- 1SpaceXAI has introduced Grok 4.7, a new flagship model optimized for complex coding and knowledge work through an enhanced reinforcement learning process and larger base architecture.
- 2The model demonstrates industry-leading price-performance metrics, particularly in long-running coding tasks like CursorBench 4.0, while maintaining competitive results across various engineering, legal, and professional benchmarks.
- 3Featuring a redesigned safeguard stack, Grok 4.7 provides robust defenses against cybersecurity and biological threats while minimizing unnecessary refusals for legitimate user requests.
The release of Grok 4.7 marks a significant advancement in SpaceXAI’s model architecture, specifically engineered for complex coding and knowledge-intensive tasks. Positioned as an evolution of the Grok 4.6 predecessor, the model emphasizes enhanced reasoning, long-context management, and robust self-verification capabilities.
Core Methodology and Technical Improvements
Grok 4.7 is built upon a larger base model architecture compared to previous iterations. The core training methodology incorporates a prolonged reinforcement learning (RL) phase, optimized through a curated dataset of multi-hour, complex tasks. This approach enables the model to effectively navigate the challenges of long-running coding projects and sophisticated reasoning chains.A critical technical feature is the native integration of the "Grok Bot harness," which aligns the model’s conversational and general knowledge workflows with its specialized logic engine. The model demonstrates superior performance in self-verification—the ability to identify and rectify errors during the generation process—which is validated by its strong standing in the CursorBench 4.0 benchmark. This benchmark specifically measures the efficacy of models in handling long-form, multi-step software engineering tasks.
Performance and Benchmarking
Grok 4.7 exhibits competitive price-performance ratios across a diverse set of professional domains:- Software Engineering: Achieving a 46.3% score on CursorBench 4.0 and 71.0% on DeepSWE v1.1, the model is highly optimized for complex codebases.
- Engineering and Professional Tasks: The model demonstrates notable proficiency in electrical engineering (EEBench: 64.0%) and professional knowledge work (GDPval Elo: 1695).
- Safety and Cybersecurity: Grok 4.7 features a redesigned safeguard stack, leading to high resistance against jailbreaking and unauthorized use. It achieves a 62.4% score on the LatchBio biosafety benchmark and limits risky prompts to a 3.3% allowance rate on HackerBench v0.3, while maintaining low false-positive refusal rates for legitimate security research.
Economic and Operational Model
The model is served through a dual-tier pricing structure:- Standard Version: Priced at \2\$6$ per million output tokens.
- Fast Variant: Offers double the output speed at double the standard token cost ( per million input, per million output).
This pricing strategy aims to provide a frontier-level model at roughly half the cost of comparable competitors. Availability is broad, spanning Cursor integration, the Grok Build platform, and direct API access for enterprise and development environments.