xAI Ships Grok 4.6 Today: Improved Post-Training on the 1.5T Core, Grok 4.7 (2.1T) Next

xAI ships Grok 4.6 today — same 1.5T V9 foundation as Grok 4.5 but with significantly better post-training (SFT + RL); Grok 4.7 at 2.1T follows in weeks.

Grok 4.6 doesn't increase parameter count — it reuses the 1.5T V9 core that xAI co-built with Cursor since that acquisition and delivers gains through improved supervised fine-tuning and reinforcement learning. API access typically follows X/SuperGrok consumer rollout by a few days. Independent benchmark scores (SWE-bench, LiveCodeBench) usually appear within a week of consumer launch.

Grok 4.7, a 2.1T model, is slated to follow Grok 4.6 by a few weeks and is projected to outperform it on every benchmark while trading slightly more latency for better token efficiency.

The practical context: the AI coding space now has four serious competitors — Claude Code, OpenAI Codex/GPT-5.6, Meta Muse Code (launched Aug 5), and xAI Grok — each shipping faster post-training iteration cycles than new architectures. For developers building model-agnostic tools, OpenRouter and LiteLLM routing matters more than model loyalty.

Why it matters: Four competitive AI coding agents on rapid-release cycles means benchmark gaps close within weeks; routing flexibility beats picking a permanent winner.

Sources: Grok 4.6 release confirmed — AIToolsReview · Grok 4.6 specs — kie.ai · Grok 4.6 & 4.7 roadmap — tbreak.com