Grok 4 Fast

Launched September 19, 2025, Grok 4 Fast is xAI's efficiency play: a single model that unifies reasoning and non-reasoning modes instead of shipping them as separate weights. It was the first xAI model with a 2-million-token context window, and xAI trained it with large-scale reinforcement learning aimed at intelligence density — the company claims roughly 40% fewer thinking tokens than Grok 4 for equivalent benchmark performance, translating to about a 98% lower cost per task. It shipped in two API variants, grok-4-fast-reasoning and grok-4-fast-non-reasoning.

The model is explicitly built for agentic use: native tool use, web and X search, and code execution are first-class, and it topped LMArena's Search Arena leaderboard at launch. Artificial Analysis independently rated it among the best price-to-intelligence ratios of any frontier model, at $0.20 per million input tokens and $0.50 per million output tokens.

Limitations: it is API-only with no open weights, so there is no self-hosting path. One third-party hallucination audit later reported a regression versus Grok 4 on its hallucination metric, and pricing above 128K tokens rises — verify current tiered pricing before long-context workloads.

92.0% AIME 2025, 93.3% HMMT 2025, 85.7% GPQA Diamond, 80.0% LiveCodeBench (xAI-reported); #1 LMArena Search Arena at launch