xAI Releases Grok 4.6: Focused on Long-Running Agents, a Clear Jump Over the Previous Version
xAI (branded on the page as SpaceXAI) released its new model, Grok 4.6. The focus is long-running agents and more ambitious visual/interactive work, and it's said to be better at testing and verifying its own output.
In brief
- What shipped: Grok 4.6, on August 12. It improves clearly over the previous 4.5 across every shared benchmark. xAI says it's especially strong at "turning a broad product idea into a working first version."
- Price: $2 input / $6 output per million tokens (the fast variant costs twice as much). That's cheap for a model in this class.
- Comparison (xAI's own measurements): a neck-and-neck picture against GPT-5.6 Sol and Fable 5 — ahead on some (e.g., GDPVal-AA, AA-Briefcase, Harvey LAB, and above GPT on CursorBench), behind on others (below both on Terminal-Bench 3.0 and DeepSWE). So it's competitive with the frontier, but not "first at everything."
- Access: through Cursor and Grok Build, via the xAI API, and through partners like OpenRouter, Vercel, and Cloudflare. Grok Build and Cursor include 2x usage for the first week.
Our take
Two things are true at once here. One is clear: the model is a real generational jump over its predecessor — it beats 4.5 on every shared benchmark. The other is more mixed: against frontier models (GPT-5.6, Fable 5) it's ahead on some tests and behind on others, notably trailing both on Terminal-Bench and DeepSWE. For our audience, the part that stands out is price: $2/$6 per million tokens is cheap for this class, and the model is widely available (Cursor, OpenRouter, Vercel, Cloudflare). So if you do coding, agent, or app-building work, it's a real, low-cost option worth trying. Two honest caveats: the numbers are xAI's own tests, and independent comparisons take weeks; and "competitive" isn't "best" — it's clearly behind on some coding benchmarks. In short: strong on price and access, but wait for independent tests before any firm verdict.
Kaynak: xAI