Grok is the model family of xAI, Elon Musk's AI company, distributed primarily through the X platform. The line is defined less by architectural novelty than by the pace of its compute scaling on the Colossus cluster and by a marketed lighter-touch refusal posture that has produced both a devoted user base and the field's most public safety failures.1
Family tree
Generations
Grok-1 shipped within months of xAI's founding, and its March 2024 open-weights release, a 314-billion-parameter mixture-of-experts under Apache 2.0, was then the largest open model drop.2 Grok-2 added image generation via partnership models; Grok 3 (February 2025) debuted alongside the claim of a 100,000-plus GPU training cluster, with frontier benchmark claims as reported by the developer; Grok 4 (July 2025) leaned into reasoning training of the verifiable-reward kind, per the developer's launch materials.3 2026 refreshes through Grok 4.5 are covered per the epistemic rules for recent releases. A planned Grok 5 had, as of the July 2026 run date, no confirmed release.1
Reception
Grok's differentiators, real-time X data and fewer refusals, cut both ways: the deployed assistant produced the July 2025 extremist-output incident and other moderation failures documented on this wiki's incident pages, while its benchmark trajectory kept xAI in the frontier conversation with OpenAI, Anthropic, and Google DeepMind on suites like MMLU and reasoning evals.1
