xAI launches Grok 3 with Think mode and DeepSearch

Background

The Grok 3 launch centred on reasoning, a capability in which a model works through a problem in steps before giving an answer.^[1]

What happened

On 17 February 2025 xAI unveiled Grok 3 and Grok 3 mini. Its announcement post followed on 19 February.^[2][1] xAI said Grok 3 was trained on the Colossus supercluster with ten times the compute of previous state-of-the-art models.^[1]

Two features came with the launch. The Think button lets users switch on reasoning trained with large-scale reinforcement learning, and it shows the reasoning so users can inspect it.^[1] DeepSearch, a research agent that synthesises information from the web into reports, launched at the same time.^[1]

Key details

xAI reported 93.3% on AIME 2025 and 84.6% on GPQA for Grok 3 (Think), and a Chatbot Arena Elo of 1402.^[1] These are the company's own figures, published with the launch.

Premium+ users got Think and DeepSearch right away. Grok 3 was made available to free users on 20 February 2025.^[1][2] That staged access meant the paid tier saw the new features first, and the free tier followed within days.

A few terms explain these features. A reasoning model spends extra computing effort working through a problem before it answers. A research agent carries out a multi-step task on its own, such as searching, reading and summarising sources. Think is the user-facing switch for reasoning, while DeepSearch is an example of an agent that works across the web. Elo is a rating system that ranks competitors from pairwise comparisons, so a higher score means the model wins more head-to-head votes.

Impact

The launch brought reasoning and a research agent into the consumer Grok product, not only into developer tools. For readers, the key point is that the Think button lets a user see the steps a model took, which makes its answers easier to check.