xAI released Grok 4.6 on 12 August, describing it as a post-training refinement of Grok 4.5 rather than a new base model. Both sit on the same 1.5-trillion-parameter "V9" foundation.
What actually changed
The work is supervised fine-tuning plus reinforcement learning in agentic environments — long-running tool use, not raw scale. xAI reports 61 on the Artificial Analysis Intelligence Index against 56 for Grok 4.5, which ties GPT-5.6 Sol on that index. Context stays at 500K tokens, inputs are text and images, and the knowledge cutoff is 1 February 2026. A new xhigh reasoning-effort tier is exposed to callers.
Where it does not lead
On coding the picture is less flattering. Grok 4.6 posts 69.9% on CursorBench v3.2, 65.9% on DeepSWE v1.1, 61.3% on FrontierCode v1.1 and 26% on Terminal-Bench v3.0. The DeepSWE number sits meaningfully below GPT-5.6's roughly 73% — a single composite index tie is not a lead on software engineering.
Pricing and availability
Below a 200K-token prompt the model costs $2 per million input tokens, $0.50 cached and $6 output; above it, $4 / $1 / $12. The "fast" variant doubles those. It is the default in Grok Build, available to all Cursor plans, and shipped with doubled included usage for the first week. There are no open weights.
A number worth deleting
Coverage has circulated a "1753 ELO" figure for Grok 4.6. That is the model's GDPVal-AA v2 score, not an Arena ELO, and the two are not comparable. A claimed rollout beginning 7 August also has no support in xAI's own materials.
The pattern this fits
A five-point index gain with no new pretraining run is the current shape of frontier competition: the expensive part is finished, and the returns are coming from what happens afterwards. It also makes version numbers a poor guide to how much compute separates two models.