Skip to main content

Grok 4.6 Arrives: SpaceX AI's New Model Matches GPT-5.6 Sol

SpaceX AI has just dropped its newest flagship model, Grok 4.6, and it's making some serious waves in the AI world. This isn't just a minor tweak—it's a major upgrade focused on handling complex, long-term tasks that require sustained reasoning and execution. And the numbers? They're impressive. In several core benchmarks, Grok 4.6 is now matching—and in some cases, even surpassing—OpenAI's leading model, GPT-5.6 Sol.

To understand how we got here, let's rewind a bit. Back in July, SpaceX AI teamed up with Cursor to release Grok 4.5, which was trained using the muscle of tens of thousands of NVIDIA GB300 GPUs. That was a big deal, but Grok 4.6 takes it a step further. The training process this time around was laser-focused on long-term complex tasks, using a ton of model-generated data that includes reasoning capabilities, advanced technical concepts, and high-quality engineering code. But that's not all—the model also went through extensive agent-based reinforcement learning, covering everything from knowledge work to general programming and specialized domain environments. The result? When you ask Grok 4.6 to build something visual or interactive, it comes up with better initial results right out of the gate.

Image

So, how does it stack up against the competition? According to the folks at Artificial Analysis, who run a comprehensive index of nine popular AI benchmarks, Grok 4.6's overall score has caught up with GPT-5.6 Sol. It's now sitting pretty behind only Claude Opus 5 and Fable 5 Max. Digging into the specifics, on the GDPval-AA v2 test—which measures logic and knowledge—Grok 4.6 scored an Elo of 1753. And on the private AA-Briefcase benchmark, which evaluates long-term knowledge work agent tasks, it hit 1577. Both of those are second only to Claude Opus 5. Not too shabby.

But performance isn't the whole story. Grok 4.6 also keeps up a competitive pace and price point. It runs at 80 tokens per second, and the pricing is pretty reasonable: $2 per million input tokens and $6 per million output tokens for the base version. If you need more speed, the high-speed variant bumps that to $4 and $12 per million tokens, respectively. And the best part? You can start using it right now if you're on Cursor or Grok Build, and developers can hook into it via the API.

So, what does this mean for the AI landscape? Well, it's clear that SpaceX AI is not just playing catch-up—they're pushing the envelope. With Grok 4.6, they've shown they can go toe-to-toe with the big names, and in some areas, even lead the pack. It's an exciting time for AI enthusiasts and developers alike, and we can't wait to see what they do next.

Key Points

  • Grok 4.6 officially released by SpaceX AI, focusing on complex, long-term agent tasks.
  • Benchmark performance matches GPT-5.6 Sol, trailing only Claude Opus 5 and Fable 5 Max.
  • Training innovations include model-generated data and extensive agent-based reinforcement learning.
  • Pricing and speed: 80 TPS, with base pricing at $2/$6 per million tokens (input/output), and high-speed at $4/$12.
  • Availability: Now live on Cursor and Grok Build, with API access for developers.