GROK 4.7: BETTER AT LONG TASKS, BUT NOT STRONGER EVERYWHERE xAI has released Grok 4.7, a model designed for coding, AI agents, and tasks that can run for hours. The new version comes with a new base model, longer RL training on difficult tasks, improved self-checking, and a long context window of up to 500K tokens. It’s already available in Cursor, Grok Build, and the API. According to xAI’s benchmarks, Grok 4.7 shows clear improvements: - CursorBench 4.0: 46.3% vs. 40.4% for Grok 4.6 - DeepSWE: 71% vs. 65.2% - Terminal-Bench: 38% vs. 20.3% On CursorBench, Grok 4.7 beat GPT-5.6 Sol, but still fell behind Fable 5.1. Independent tests paint a more mixed picture. Artificial Analysis gives Grok 4.7 a small improvement - 46 vs. 44 on its Intelligence Index. As a coding agent inside Grok Build, however, the model appears to be stronger. There’s also a downside: Grok 4.7 is more verbose. The token price remains unchanged, but longer outputs mean some tasks can end up costing more. API pricing: $2 per 1M input tokens / $6 per 1M output tokens. Once the prompt exceeds 200K tokens, the pricing rate doubles. The Fast version is 2x faster and more expensive, and is currently available only in Cursor and Grok Build. Overall, Grok 4.7 makes the most sense for long coding sessions, large repositories, and complex multi-step tasks. For short prompts, the improvement may be much smaller. So is Grok 4.7 actually a major upgrade - or just a more expensive model that shines in a narrower set of tasks?