Grok 4.5 is xAI's latest flagship, built on a 1.5 trillion parameter mixture-of-experts architecture (V9 foundation). Its standout feature is configurable reasoning — you can dial compute from low (fast, cheap) to high (deep, thorough) per request, letting a single model cover use cases from interactive chat to complex analysis. It integrates natively with X search and web search, and supports function calling and structured JSON output. At $2/$6 per million tokens with a 500K context window, it slots in as a cost-effective frontier alternative. Trained partly on Cursor session data for realistic coding workflows.
| Benchmark | Score | Type | Recorded |
|---|---|---|---|
| Humanity's Last Exam | 42.7 | accuracy | 29d ago |
| SWE-Bench | 86.6 | accuracy | 29d ago |
| GPQA Diamond | 93.1 | accuracy | 29d ago |
| LiveCodeBench | 87.4 | accuracy | 29d ago |
| SciCode | 55.0 | accuracy | 29d ago |
| MMLU-Pro | 89.2 | accuracy | 29d ago |
| LCR | 79.3 | accuracy | 29d ago |