Released on 12 August 2026, Grok 4.6 replaces Grok 4.5 as xAI's frontier model and is aimed squarely at work that runs for many steps: multi-stage research, changes spread across a codebase, and turning a product idea into a working first version. xAI reports more self-testing on longer trajectories, with the model verifying its own output before continuing.
On xAI's own comparison table it scores 61 on the Artificial Analysis Intelligence Index, level with GPT-5.6 Sol Max and just under Claude Fable 5 at 62, and leads the field on CursorBench v3.2 (69.9%) and on the Harvey LAB legal evaluation (15.8%). It gives ground elsewhere: DeepSWE v1.1 65.9% against 73% for GPT-5.6 Sol Max, and Terminal-Bench v3.0 26% against 34.6%. Compared with Grok 4.5, the gains are consistent rather than uniform — GDPVal-AA v2 rises from 1526 to 1753 and APEX-Agents from 47.1% to 57.5%.
The context window stays at 500,000 tokens and the headline rate is unchanged from Grok 4.5, but cached input is more expensive: $0.50 per million against $0.30. The model launched in Cursor and in xAI's own Grok Build environment, with doubled included usage in both for the first week. Its knowledge cut-off is 1 February 2026.