Grok 4.7 Reaches Frontier on Independent Tests, But Uses 3x GPT-6 Astra's Tokens
Artificial Analysis released its independent benchmarking of Grok 4.7 on September 21, providing the first third-party evaluation of xAI’s latest model beyond the company’s own announcement. The headline number is an Intelligence Index score of 46 — placing Grok 4.7 two points above Grok 4.6 and bringing xAI into the top four AI labs by Artificial Analysis’s composite assessment.
The gains are genuine. The cost to produce them is significant.
What the Numbers Show
On AA-Briefcase, Artificial Analysis’s private benchmark for long-horizon agentic knowledge work, Grok 4.7 scores 1657 ELO at xhigh effort — up 111 points from Grok 4.6 (high) at 1546. That places it just behind Claude Opus 5 and Claude Fable 5.1 at the frontier.
On GDPval-AA, which evaluates models producing practical work outputs like documents, spreadsheets, and slides, Grok 4.7 scores 1695 ELO versus 1605 for Grok 4.6 (high).
The coding agent performance is the most striking improvement. Grok Build with Grok 4.7 at xhigh effort scores 56 on the Artificial Analysis Coding Agent Index, up from 47 with Grok 4.6 — a 9-point gain. Among models running in their native harnesses, Grok 4.7 now ranks fourth behind Claude Fable 5.1, GPT-6 Astra, and Claude Opus 5. Notably, it has overtaken GPT-5.6 Sol on the coding agent ranking.
The Token Efficiency Problem
The gains come at a price measured in tokens. Grok 4.7 at xhigh effort uses 81,000 output tokens per task. That is more than double Grok 4.6’s 36,000 tokens at high effort, and three times GPT-6 Astra’s 27,000.
For comparison:
| Model | Output Tokens / Task |
|---|---|
| GPT-6 Astra (max) | 27,000 |
| Grok 4.6 (high) | 36,000 |
| Muse Spark 1.3 (max) | 60,000 |
| Claude Fable 5.1 (max) | ~78,000 |
| Grok 4.7 (xhigh) | 81,000 |
At frontier output pricing, a 3x token gap versus GPT-6 Astra has real cost implications for any production agentic workload. If the quality gap were proportional, the trade-off might be obvious. It is not. GPT-6 Astra still outranks Grok 4.7 in the coding agent standings.
Reading the Gains
The benchmarks suggest Grok 4.7’s improvements come from extended reasoning — the model is spending more compute per task and achieving better results on long-horizon work as a consequence. xAI’s training emphasis on multi-hour tasks and self-verification shows up clearly in the AA-Briefcase and GDPval scores, where sustained, structured effort matters.
That is a defensible design choice for models targeting deep knowledge work. For coding agent use cases where developers run hundreds of tasks daily, the token count will be a limiting factor in deployment decisions. Grok 4.7 is unambiguously a better model than Grok 4.6. Whether it is the right choice at three times the token output of its nearest competitor depends entirely on what the task is.