What's this about?
SpaceXAI released Grok 4.6, the newest version of its Grok model. On the industry's most-watched composite score — the Artificial Analysis Intelligence Index — it scores 61, which ties it with OpenAI's GPT-5.6 Sol and puts it just behind Claude Opus 5 (63) and Fable 5 (62). What makes it notable isn't the raw score but the price: it costs $2 per million input tokens and $6 per million output tokens, more than 60% less than Opus 5 or GPT-5.6 Sol, while performing at roughly the same level on agent tasks.
What exactly is an "Intelligence Index"?
No single test can tell you how smart a model is, so Artificial Analysis (an independent benchmarking company) runs nine different tests — coding, reasoning, knowledge, agent tasks — and averages them into one number out of 100. Think of it like a decathlon score rather than a single race time: a model can't win by being great at one narrow thing. Because it's run by an outsider rather than the model's maker, people trust it more than a lab's own published charts.
What's the community saying?
- Price is the headline, not the benchmark. The most-upvoted reaction was some version of "Fable-level performance, faster and significantly cheaper" — the consensus is that Grok 4.6 doesn't beat the best models, it makes their performance level much cheaper to buy.
- The Cursor connection. Several commenters pointed out that SpaceXAI's acquisition of Cursor now looks strategically shrewd — one argued Musk "made a savvy choice buying Cursor's data," and others noted Cursor subscriptions now stretch much further than OpenAI's or Anthropic's because Grok is a first-party model there.
- Cache pricing quietly went up. A heavy user flagged that cache-read pricing nearly doubled from $0.30 (Grok 4.5) to $0.50, which matters because in long coding sessions cache reads can be ~80% of the actual bill — so the headline "unchanged pricing" isn't the whole story.
- Vertical integration as the long game. One popular comment argued SpaceXAI is the only frontier lab that owns its compute, data centers, and soon chip fabrication, predicting it pulls ahead on cost per token even if not on raw intelligence.
- Political discomfort persists. A visible minority of comments refused to evaluate the model on technical merits at all, and others said they had never met anyone who actually uses Grok for coding — brand remains a real adoption drag regardless of benchmarks.