xAI ships Grok 4.6, a post-training-only upgrade it says matches GPT-5.6 Sol

xAI released Grok 4.6 on August 12, 2026, keeping the same 1.5-trillion-parameter foundation as its predecessor Grok 4.5 and routing the entire upgrade through post-training rather than a new base model.

The company says the changes — an extended supervised fine-tuning and reinforcement-learning run on curated reasoning, engineering and web-development data, plus a revised optimizer — improved every benchmark it reports over Grok 4.5. That includes an Artificial Analysis Intelligence Index score of 61, which xAI says matches rival GPT-5.6 Sol, a CursorBench 3.2 score of 69.9%, and a GDPVal-AA score of 1753.

The approach matters because it points to headroom without the cost of training a new foundation model: xAI is extracting gains from data and tuning on the same 1.5-trillion-parameter V9 base.

Grok 4.6 costs $2 per million input tokens and $6 per million output tokens, with a faster variant priced at double. It is available in the Cursor editor, Grok Build, the xAI application programming interface and partner platforms, with double the included usage in Grok Build and Cursor for the first week.

The benchmark table is xAI’s own. The Artificial Analysis Intelligence Index is a third-party composite, but the reported scores are self-supplied and not independently verified, and matching a rival on one aggregate index does not establish parity across coding, agentic or long-context work.