Notes ·
Grok 4.6 Makes My Switch Away from Claude Look Less Unusual
Artificial Analysis has published its first results for Grok 4.6, and the interesting story is not simply that another model moved up a leaderboard.
Grok 4.6 scores 61 on the Artificial Analysis Intelligence Index, equal to GPT-5.6 Sol and just behind Claude Opus 5 at 63 and Claude Fable 5 at 62. More importantly for coding agents, it scores 88.4% on Terminal-Bench v2.1 and sits near the top of Artificial Analysis's agentic evaluations.
It is also remarkably efficient. On Artificial Analysis's long-horizon knowledge-work benchmark, Grok 4.6 completed tasks in roughly 53 turns, compared with about 103 turns for Claude Opus 5, while consuming about one-quarter as many input tokens. Its measured cost was $0.84 per task.
That matches something I have already noticed using Grok 4.5 with Grok Build for my personal coding.
I moved away from Claude because Grok communicates with me better.
That sounds like a small thing. It isn't.
I don't want an agent giving me a wall of prose or creating another pile of work for me to interpret. I want it to tell me what I need to know, do the work we agreed on and let me make the decisions.
Grok is fast enough that the interaction feels much more like pair programming. I steer, it works, we look at the result and continue. That has improved both my productivity and my output.
Grok Build itself is also now open source, with support for planning, subagents, skills, MCP, hooks, code review and local models. Grok 4.5 remains the model officially documented as powering it today.
I still cross-check difficult problems with Codex or GPT-5.6 Sol when another perspective is useful. But Grok has become my daily driver.
Anthropic's newly announced invisible watermarking of Claude-generated content gives me another reason to prefer alternatives for my own work.
Benchmarks do not measure whether a model communicates in the way a particular developer prefers. But the new results suggest the speed and efficiency I have been experiencing are not imaginary.
If you're curious, install Grok Build and give it a real coding task.
You might be surprised.