SpaceXAI announced on September 21 that it has released Grok 4.7, a new AI model designed to be its most powerful option for coding and knowledge-intensive work. The model utilizes a larger base model than its predecessor, Grok 4.6, and has undergone extended reinforcement learning focused on complex, long-duration tasks. These updates have reportedly enhanced the model's ability to verify its own work and manage longer context.
Model ReleasesSpaceX AIGrok 4.7
SpaceXAI Unveils Grok 4.7 with Enhanced Coding and Knowledge Work Performance
In benchmark testing, Grok 4.7 demonstrated significant performance gains in specific areas. In the CursorBench 4.0 benchmark for long-duration coding tasks, Grok 4.7 achieved a score of 46.3%, outperforming Grok 4.6 (40.4%) and OpenAI's GPT-5.6 Sol (41.7%), though it trailed behind Anthropic's Claude Fable 5.1 (51.8%). It also ranked highest among tested models in the EEBench (64.0%) and the Harvey Legal Agent Benchmark (19.6%) for legal tasks.
The model features a completely redesigned safeguard stack to improve refusal and resistance to jailbreaking. In the HackerBench v0.3 benchmark for malicious cyber tasks, the model allowed only 3.3% of risky dual-use prompts through, while reportedly minimizing blocks on legitimate security work.
Grok 4.7 is available immediately through Cursor, Grok Build, and the Grok API. API pricing starts at $2 per million input tokens and $6 per million output tokens. The company also offers a high-speed variant that doubles the output rate at twice the cost.
Sources
- SpaceXAI、「Grok 4.7」発表 コーディングと知識労働向けで同社最高性能、価格は据え置き （ITmedia AI+, 2026-09-22）
- 公式ブログ