Grok 4.7: The Expensive Stepchild of the AI Model Race

The claim. Elon Musk promised Grok 4.7 would exceed all current models, citing SpaceX's unique training corpus. The reality is more muddled. The model scores lower than Grok 4.6 on several benchmarks, contradicts earlier efficiency promises, and costs significantly more to run. The creator of this video argues the release is a classic case of benchmark-data disconnect, where the model's real-world behavior doesn't match the numbers being marketed.

The contradiction. Elon said 4.7 would be slower but more token-efficient. Instead, it uses roughly 2x the tokens of 4.6 in benchmarks like Artificial Analysis, driving costs above GPT-4.5 Astra. When Cursor's CEO Michael Truel claimed production usage showed only 5-30% more tokens per request, the video creator counters that requests per prompt have increased, negating any efficiency gain. The claim either misrepresents metrics or is misleading.

The messy release. The video highlights erratic behavior: the model fell into endless loops in Grok Build, generated terrible front-end designs, leaked raw markdown system prompts on X.com, and took 90 minutes for a simple game demo. SpaceX appears to have rushed an RL-fix during the last week, making the model more verbose but less efficient. This feels like a stepping stone, not the generational leap promised.

The real strength. Despite the noise, Grok 4.7 has higher floor quality than peers. In a custom benchmark analyzing a codebase for improvements, it performed just behind GPT-6 Astra on finding viable suggestions, and ahead of Claude 5.1 Fable. It stays on task longer, checks its work more thoroughly, and catches things other models miss. For teams that value breadth and depth over speed, this matters.

The cost problem. The capability jump does not justify the cost increase. In the video's benchmark, best-case runs matched Astra and Fable, worst-case cost twice as much. Subscription plans now cap users aggressively: $40 of usage burned 8% of a weekly limit on the $300 plan. Compared to the near-unlimited usage of Grok 4.5, this is a severe regression. There is no economic reason to switch.

The verdict. Grok 4.7 is a fine model for curious tinkerers, especially those already in the Cursor ecosystem. But for engineering leaders evaluating tooling, it's hard to recommend. The cost, inconsistent behavior, and marketing misalignment make it a wait-and-see release. The next iteration might deliver on promises, but this one is a flop at current pricing and reliability.