Grok 4.5: ~4.2x fewer output tokens than Opus 4.8 max on SWE Bench Pro. 80 TPS serving. #1 on Harvey Legal Agent Benchmark. Default in Grok Build. Trained on GB300 cluster in Memphis. Fable max leads benchmarks; Grok 4.5 closest on Terminal Bench 2.1.