AIBRIEF
Cursor, the AI coding editor many developers already use, fed real world coding sessions into xAI's post training run. Grok 4.6 is the second consecutive model to lean on that pipeline.
xAI has now shipped two Grok models in a row whose post-training leaned on Cursor, the AI coding editor many developers already use. Grok 4.6, released August 12, 2026, is the first time that bet shows up in independent benchmark numbers.
The model scores 61 on the Artificial Analysis Intelligence Index, a composite of nine standard tests, putting it in line with GPT-5.6, OpenAI's flagship, and five points above Grok 4.5. It also lands a GDPval-AA v2 Elo of 1753 (second only to Claude Opus 5), 50.7% on τ³-Banking, and 88.4% on Terminal-Bench v2.1.
xAI's pitch is that Cursor's real-world coding sessions, the ones developers produce while working in the editor, are a cleaner training signal than synthetic code. Grok 4.5 was the first xAI model trained on that data; 4.6 was trained on more of it, with a longer supplemental run. The honest counter: the index averages nine benchmarks, and the +5 gain over 4.5 could reflect the extra compute as much as the Cursor pipeline.
Elon Musk called the model 'objectively #1' on X. The benchmarks describe parity, not dominance. Grok 4.6 also arrives still carrying the non-consensual nude-image and MechaHitler episodes in its reputation, baggage the company has yet to shake. The Cursor relationship itself is described in coverage as a partnership and possibly a pending acquisition; xAI has not confirmed a deal.
Grok 4.6 ships first inside Cursor and xAI's own Grok Build, with double usage for the first week. Grok Bot, an always-on agent, entered beta the day before.