Grok 4

Grok 4 is xAI's flagship language model, trained with reinforcement learning to use tools like a code interpreter and web browser, and wired into real-time search across X and the web. It is a strong pick for deep research questions, coding, and hard multi-step analysis, with a 256K-token context window and a multi-agent Heavy variant for the toughest problems. The weights are closed and pricing sits at the high end of the market, and newer Grok versions (4.1 through 4.7) have since taken the spotlight, so it reads as a proven older flagship rather than the cutting edge.

Provider-published (xAI launch materials): Humanity's Last Exam - Grok 4 Heavy 50.7% on the text-only subset (first model above 50%); 44.4% with tools vs 25.4% without on the full set; ARC-AGI-2 - 15.9% (closed-model SOTA, nearly double Claude Opus at ~8.6%); USAMO 2025 - 61.9% (Grok 4 Heavy); GPQA - 87.5% (88.4% Heavy); Vending-Bench (agentic) - $4,694.15 net worth, topping Claude Opus 4 and human operators.