Whoa, Grok 4.7 is here, and it’s only been a little over a month since Grok 4.6. The new model switched to a larger base model and extended training specifically to tackle tasks that require working for several hours straight. Major improvements were made in coding, long-context understanding, self-checking, documentation, and agent tasks. The most insane upgrade this time is in Terminal-Bench, jumping from 20.3% straight to 38.0%, nearly doubling. CursorBench went from 40.4% → 46.3%, and DeepSWE improved from 65.2% → 71.0%. Plus, it’s starting to show some weird “specialized advantages”: Legal tasks at 19.6%, while GPT-5.6 Sol is only 2.5% and Fable 5.1 is 6.7%. Electrical engineering at 64.0%, also outperforming both competitors. However, coding isn’t universally the best—DeepSWE is still slightly below GPT-5.6 Sol, and Terminal-Bench is noticeably lower than Fable 5.1. But honestly, the craziest part is the price. Grok 4.7 switched to a larger model, with capabilities improving across the board, yet the API pricing remains at $2/M input and $6/M output, exactly the same as Grok 4.6. In SpaceXAI’s own comparison, GPT-5.6 Sol is $4/$20, and Fable 5.1 is $10/$50.