何が起こったのか
SpaceXAI has released Grok 4.7, a new flagship AI model designed for coding, agentic tasks, and knowledge work. The model is built on a larger base architecture and underwent a longer training phase compared to Grok 4.6. Despite these architectural changes, the company has maintained the same pricing structure of $2 per million input tokens and $6 per million output tokens.
SpaceXAI's Grok 4.7 is now available via the xAI API, Cursor, Grok Build, OpenRouter, Vercel, and Cloudflare. The model demonstrates performance gains over Grok 4.6, with SpaceXAI reporting a jump in Terminal-Bench 4.0 scores from 20.3% to 38.0% and an 11-point increase in EEBench to 64.0%.
The release includes a 'Grok 4.7 Fast' version, which provides twice the output speed at twice the cost. This variant is limited to Cursor and Grok Build and is excluded from the free tier of Grok Build. A US-specific API endpoint is also available at a 10% price premium.
Safety features have been updated with a new safeguard stack. SpaceXAI claims the model achieved a 62.4% score on LatchBio’s biosafety benchmark and reported a 3.3% failure rate on its internal HackerBench v0.3 for risky dual-use prompts. Select cybersecurity partners have been granted invite-only access to the model's red-teaming capabilities.
なぜそれが重要なのか
Grok 4.7 represents a significant iteration in SpaceXAI's model lineup, specifically targeting performance improvements in technical benchmarks like Terminal-Bench 4.0 and EEBench. By maintaining price parity with the previous version while increasing performance, the model aims to improve the cost-efficiency of agentic workflows. Furthermore, the introduction of a new 'safeguard stack' and specific focus on biosafety and cybersecurity red-teaming indicates a strategic shift toward enterprise-grade safety requirements. The model's availability across multiple platforms, including Cursor, Vercel, and Cloudflare, facilitates immediate integration for developers.
The model's competitive positioning is highlighted by its price-performance ratio compared to larger models like Fable 5.1 Max and GPT-5.6 Sol Max. While Fable 5.1 Max and GPT-5.6 Sol Max lead in specific benchmarks, they carry significantly higher input and output costs, making Grok 4.7 a potentially more economical choice for high-volume coding and agentic tasks.
The focus on cybersecurity and biosafety benchmarks suggests that SpaceXAI is attempting to address enterprise concerns regarding the deployment of powerful AI agents in sensitive environments. By providing tools for defense research and implementing stricter refusal mechanisms, the company is positioning Grok 4.7 as a safer alternative for professional knowledge work.
インタラクティブなメカニズム: 実際にどのように機能するか
この開発の背後にある基盤となるテクノロジーをインタラクティブに探索します。
What is the best response when AI Models Explained makes a mistake in production?
次に見るべきもの
Users should monitor the performance of the new 'Grok 4.7 Fast' variant, which offers double the output speed at double the cost but is currently restricted to Cursor and Grok Build. Additionally, the efficacy of the new safeguard stack in real-world scenarios remains to be independently verified, as current safety claims are based on vendor-reported benchmarks. The impact of the 10% premium for the US-based regional API endpoint on enterprise adoption patterns is also a key area for observation.
The reliance on vendor-reported benchmarks means that independent validation of the model's capabilities in coding and legal reasoning is necessary to confirm its standing against competitors like Fable and GPT-5.6.
The long-term stability and reliability of the new US regional endpoint and the effectiveness of the recommended 'prompt_cache_key' for developers will be critical for enterprise users requiring consistent inference performance.
The restriction of the 'Fast' variant to specific platforms may limit its utility for developers who rely on the public xAI API for their infrastructure.