xAI releases Grok 4.1 and Grok 4.1 Fast

Background

xAI's model releases in 2025 came at a fast pace, and Grok 4.1 was one of several updates in the Grok line after Grok 4. Before the official announcement, xAI tested the model on live traffic.^[1]

What happened

xAI announced Grok 4.1 on 17 November 2025. The model had been rolled out silently to a growing share of traffic from 1 to 14 November, so the announcement came after two weeks of real use.^[1] Its thinking mode took first place on the LMArena Text Arena with 1483 Elo.^[1]

Grok 4.1 was made available to all users on grok.com, on X and in the iOS and Android apps.^[1] Two days later, on 19 November 2025, xAI followed with Grok 4.1 Fast, a 2-million-token tool-calling model, together with the Agent Tools API.^[2]

Key details

The Agent Tools API runs search, file search, code execution and MCP tools on xAI's servers.^[2] MCP, the Model Context Protocol, is a standard that lets AI models connect to outside tools and data sources. Running these tools on xAI's side means the developer does not need to build and host each one separately.

xAI reported a 100% score for Grok 4.1 Fast on τ²-bench Telecom, a test of how well an agent handles telecom customer-service tasks.^[2] A perfect score on one test is a claim worth checking against independent results, and it is one of the figures the company chose to highlight.

A thinking model spends extra steps working through a question before it answers, while a non-thinking model responds directly. Offering both lets a product choose between speed and deliberation for each request. Arena rankings come from anonymous side-by-side votes, so they measure how people judged the answers rather than how the model scores on a fixed set of questions.

Impact

The release put a thinking model at the top of a public arena and paired a fast, tool-calling variant with a set of hosted tools for developers. For readers, the event shows two directions in xAI's work at once: general chat quality, and the tools that let models act on a user's behalf.