Grok 3 and Grok 3 mini

What it is

Grok 3 is a large language model from xAI, unveiled on 17 February 2025 and released as Grok 3 Beta alongside a smaller model called Grok 3 mini.^[2][1] xAI's announcement post followed on 19 February 2025.^[1] The model is built for reasoning, meaning it works through a problem step by step before answering, and for agentic tasks, where the model uses tools such as web search to complete a multi-step job.^[1]

Origins

xAI said Grok 3 was trained on Colossus, its supercluster, with ten times the compute of previous state-of-the-art models.^[1] Reports put the cluster at about 200,000 GPUs.^[2] A GPU, or graphics processing unit, is the kind of chip most large AI models are trained on, and a supercluster links very many of them to work on one training run.

Grok 3 followed Grok 2 in the Grok line, and xAI's launch centred on reasoning.^[1]

Key features

The first feature is the context window. Grok 3 can hold 1 million tokens at once, which xAI put at about eight times that of earlier Grok models.^[1] A token is a small piece of text, and a context window sets how much text the model can keep in view while it answers.

The second feature is Think mode. Its reasoning was refined through large-scale reinforcement learning, a training method in which the model improves by being rewarded for good results. In Think mode, the model thinks for seconds to minutes before responding, and users can read the reasoning it produced.^[1] This made it possible to see how the model reached an answer, not only the answer itself.

Grok 3 mini is a separate and cheaper reasoning model. xAI aimed it at STEM tasks, meaning science, technology, engineering and mathematics, that need less general world knowledge.^[1] It is a cheaper option than the full model.

The third feature is DeepSearch, a research agent launched at the same time. It searches the web, weighs conflicting sources and writes a summary report.^[1] DeepSearch is an example of an agent: the model does not only answer from what it already knows, but gathers new material and organises it for the user.

Benchmark results

xAI reported several results for Grok 3. Grok 3 (Think) scored 93.3% on AIME 2025, a set of maths competition problems, at its highest test-time compute, which means the amount of computation spent on each answer. It scored 84.6% on GPQA, a set of graduate-level science questions, and 79.4% on LiveCodeBench, a test of programming problems.^[1]

Grok 3 mini scored 95.8% on AIME 2024 and 80.4% on LiveCodeBench, according to xAI.^[1] The non-reasoning version of Grok 3 scored 73.2% on MMMU, a test that asks questions about images that require college-level knowledge.^[1] On the Chatbot Arena, a site where users compare anonymous chatbots, Grok 3 reached an Elo score of 1402.^[1] Benchmark figures reported by the developer show one view of performance, so they are best read as the developer's own claims.

Availability

Grok 3 was offered to X Premium and Premium+ users and on Grok.com. Free users got access on 20 February 2025.^[1][2]

The API launched in April 2025 at $3 per million input tokens and $15 per million output tokens.^[3][2] In May 2025 Grok 3 became available on Microsoft Azure.^[3][2]

Place in the line

Grok 3 introduced the Think button and DeepSearch to the Grok line. It was succeeded by Grok 4 in July 2025.^[2]