Grok 4.6 Ships Without a Model Card or Benchmarks

xAI shipped Grok 4.6 on August 7 with no model card, no benchmark scores, and no way to verify what changed, per reporting from Evolink.ai. That leaves developers guessing: every performance number attached to 4.6 is speculation, not measurement, until independent Arena testing produces a score.

The model reuses the same 1.5 trillion-parameter V9 foundation that powered Grok 4.5. xAI is crediting the gains to improved supervised fine-tuning and reinforcement learning, not a bigger base model, according to the same report. That's a meaningful distinction if true: whatever 4.6 does better, it's doing with the same raw compute as its predecessor.

For context, Grok 4.5 in its "high" configuration scores 54 on the Artificial Analysis Intelligence Index, putting it fourth behind Claude Fable 5, GPT-5.5, and Claude Opus 4.8. Whether 4.6 moves that number up, down, or stays flat is unknown until someone runs the tests xAI didn't publish.

A 2.1T-parameter Grok 4.7 is reportedly weeks away, per the same August 7 report.

Why skip the paperwork

Shipping a mid-cycle model with zero verification while a much bigger successor is already queued reads like a placeholder move, not a flagship claim. If 4.6 mattered as a capability jump, xAI had every reason to publish the numbers on release day. Watch whether Arena scores show up before 4.7 lands. If they don't, 4.6 was never meant to be judged on its own.