xAI Launches Grok 4.6 — Better Reasoning, Same Scale, Beats Claude Opus 4.8 on Coding
Summary: xAI released Grok 4.6 on August 7, delivering measurable coding gains over Grok 4.5 without increasing model size — a bet on training quality over raw scale.
Key Facts
- Architecture unchanged at 1.5T parameters (V9); improvements come entirely from enhanced SFT and RL training recipes
- SWE-Marathon score: 29.0%, clearing Claude Opus 4.8's 26.0% and positioning Grok as a competitive coding agent
- Throughput at 80 transactions per second; pricing held steady at Grok 4.5 levels
- Grok 4.7 — a substantially larger 2.1T-parameter model — confirmed for release within weeks
Why It Matters
Grok 4.6 is a case study in the shift from compute-driven scaling to training-driven refinement. As frontier labs face diminishing returns on raw parameter counts, the ability to extract more capability per training run is becoming a key differentiator. The cadence also signals xAI's strategy: rapid iterative releases rather than occasional large leaps — putting pressure on rivals to match both quality and update speed.