For years, frontier AI pricing followed a familiar restaurant rule: the fancier the menu, the more carefully you watched the bill. Grok 4.6 breaks that pattern. It scores 61 on the Artificial Analysis Intelligence Index, level with GPT-5.6 Sol and only behind the highest Opus 5 and Fable 5 configurations, while keeping the same $2/$6 per-million-token price as Grok 4.5.
The more interesting story sits below the composite score. On GDPVal-AA v2, which tests professional agent work, Grok 4.6 reaches 1753 Elo. On AA-Briefcase, a long-horizon knowledge-work evaluation, it reaches 1577 Elo. Artificial Analysis found that it completed those tasks in roughly 53 turns and 0.5 billion input tokens on average, compared with about 103 turns and 2.0 billion for Opus 5 max. Imagine two teams reaching a similar report: one holds half as many meetings and rereads a quarter as much paperwork. That efficiency changes both the invoice and the chance that a long agent run wanders away from its goal.
xAI trained directly for this kind of sustained work. The launch describes longer supplemental training, regenerated reasoning and software trajectories, and reinforcement learning across knowledge work, general coding, web development, CAD, and other tool environments. The model has a 500K-token context window and low, medium, high, and xhigh reasoning levels. It is available in Grok Build, Cursor, the xAI API, OpenRouter, Vercel, and Cloudflare; a faster API variant costs twice the standard rate.
The release is not a clean sweep. Grok’s strongest evidence is agentic knowledge work. On xAI’s table, DeepSWE v1.1 at 65.9% trails GPT-5.6 Sol and Fable 5, while Terminal-Bench v3.0 at 26% trails both by a wider margin. Some competitor values come from different cards or leaderboards, so tiny gaps should not be treated as laboratory precision. Early user reports also include unsafe security edits and poor behavior after failures. Those reports are anecdotal, but the correct response is simple: sandbox the agent, minimize credentials, inspect diffs, and run tests.
The practical conclusion is not “trust Grok more.” It is “you can afford to verify Grok more.” A measured $0.84 per task makes a second run, source check, or reviewer model easier to justify. For research, analysis, office artifacts, and long agent jobs, Grok 4.6 is now a genuine frontier option with a value advantage. For high-stakes facts or production access, keep a human at the gate.
