Grok 3 is xAI’s model released on February 17, 2025, trained on the Colossus supercomputer in Memphis using more than 200,000 GPUs, roughly ten times the compute xAI put into Grok-2. xAI pitched it as a reasoning model with a “Think” mode that can spend anywhere from seconds to minutes working through a problem, checking its own steps and backtracking before answering, and it reported an Elo score of 1402 on Chatbot Arena along with strong results on GPQA, AIME math competition problems, and MMLU-Pro. The context window grew to 1 million tokens, eight times what earlier Grok models supported, letting it handle much longer documents and codebases in one request. Grok 3 is priced at $3 per million input tokens and $15 per million output tokens through xAI’s API, and it’s also available through the Grok app and X. xAI positioned it against OpenAI’s o3, Gemini 2 Pro, and DeepSeek V3, claiming an edge in math and science reasoning at launch.