xAI：News（网页）

精选

xAI发布Grok 4.1 Fast与Agent Tools API

2025-11-19 08:00· 306天前

AI 导读

xAI发布Grok 4.1 Fast模型及Agent Tools API。Grok 4.1 Fast支持200万token上下文，在τ²-bench Telecom基准测试中获100%得分且成本仅105美元，函数调用准确率72%。Agent Tools API集成实时X数据、网页搜索与代码执行功能。该模型在深度研究基准测试中超越GPT-5等竞品，成本更低且幻觉率较上代降低一半。

推荐理由

xAI 发布 Grok 4.1 Fast 及 Agent Tools API，支持 2M 上下文与原生工具调用，剑指企业级 Agent 应用。

原文 · 保持原样，未翻译原文 · 未翻译

Grok 4.1 Fast and Agent Tools API Trained for the real world State-of-the-art tool calling Agent Tools API A full-featured toolset The best agent for deep research Start Building

Booking Agent

You are a hotel customer support agent, you can help guests make and manage bookings and general enquiries.

Available tools

client-tool Find guest by email

server-code Get booking details

web-search Find available rooms

client-tool Modify booking

Hi, I'd like to upgrade my current booking to an Executive Suite.

Sure, give me moment to find your booking...

Finding your account...

Searching available rooms...

Upgrading booking to Executive Suite...

Okay, all done. You’re now booked into the Executive Suite - enjoy your stay!

Thank you!

Plan of action

Identify guest

Check availability

Upgrade booking

client-tool Find guest by email

Input

Email:john.doe@example.com

Output

User ID:1234567890

server-code Get booking details

Input

User ID:1234567890

Date:01 Dec - 05 Dec, 2025

Output

Booking ID:BX-23929

web-search Find available rooms

Input

Date:01 Dec - 05 Dec, 2025

Room type:Executive Suite

Output

Room ID:2918482

client-tool Modify booking

Input

Booking ID:BX-23929

Room ID:2918482

Today, we’re excited to launch two powerful new additions to the xAI API:

Grok 4.1 Fast, our best tool-calling model with a 2M context window. It reasons and completes agentic tasks accurately and rapidly, excelling at complex real-world use cases such as customer support and finance.
The Agent Tools API, which gives agents access to real-time X data, web search, remote code execution, and more.

Paired together, Grok 4.1 Fast and the Agent Tools API empower developers to build production-grade agents that specialize in tool calling and agentic search.

Trained for the real world

We built Grok 4.1 Fast specifically for real-world enterprise use cases.

Through RL training in simulated environments, Grok 4.1 Fast was exposed to a wide variety of tools covering dozens of domains. This diverse training gives Grok 4.1 Fast exceptional performance on τ²-bench Telecom, a challenging benchmark that evaluates agentic tool use in real-world customer support scenarios.

τ²-bench Telecom

Score(%)

Total Cost($)

Independent evaluation verified byArtificial Analysis

State-of-the-art tool calling

As developers build increasingly capable autonomous agents that plan over long horizons and operate independently, models must deliver intelligence without compromising speed and cost.

Grok 4.1 Fast is our answer: a model that combines frontier tool-calling performance with blazing-fast inference and cost effectiveness.

Berkeley Function Calling v4 Benchmark

Overall Accuracy(%)

Total Cost($)

*Gemini 3 Pro's score is an estimate provided by an independent evaluator, pending official results.

A common challenge for agentic models is that performance degrades as context length increases. We trained Grok 4.1 Fast using long-horizon reinforcement learning with a strong emphasis on multi-turn scenarios, ensuring consistent performance across its full 2-million-token context window.

Multi Turn Acc

Multi Turn Long Context

Grok 4.1 Fast

Grok 4 Fast

Grok 4

Agent Tools API

We’re also launching the Agent Tools API, a suite of powerful server-side tools that allow Grok 4.1 Fast to operate as a fully autonomous agent.

How are people reacting to Tesla Robotaxi?

Agent Grok 4.1 Fast

Robotaxi Reactions

Strong excitement over Arizona approval and rapid expansion, with many praising cheaper and smoother rides compared to other rideshare services. Overall positive momentum among 𝕏 users and the general public.

Browsing the web for news about Robotaxi

Reading X Posts to gauge community reception

Building a sentiment chart using Python

With just a few lines of code, developers can enable Grok to browse the web, search X posts, execute code, retrieve uploaded documents, and more.

import os
from xai_sdk import Client
from xai_sdk.tools import code_execution, web_search, x_search, collections_search, mcp

client = Client(api_key=os.getenv("XAI_API_KEY"))
chat = client.chat.create(
    model="grok-4-1-fast-reasoning",
    tools=[
        web_search(),
        x_search(),
        code_execution(),
        collections_search(collection_ids=["..."]),
        mcp(server_url="..."),
    ],
)

These tools run entirely on xAI’s infrastructure, so developers no longer need to manage API keys, rate limits, sandboxes, or retrieval pipelines. Grok decides when and how to use them, often invoking multiple tools in parallel across several turns, until it has everything it needs to deliver a final answer.

A full-featured toolset

The Agent Tools API is a versatile suite that lets you significantly extend the capabilities of our base Grok models. Key features include:

Search Tools

Harness realtime X and internet search for fast, comprehensive insights into current events and trends.

Files Search

Intelligently search and retrieve relevant documents from your uploaded files, with citations.

Code Execution

Execute Python code in a secure sandbox to analyze data and run simulations.

MCP Tools

Connect seamlessly to external MCP servers, enabling access to powerful custom third-party tools.

The best agent for deep research

Real-time information retrieval and deep research are core strengths of Grok 4.1 Fast. With our native integration into the X ecosystem and powerful web-browsing capabilities, search agents powered by the xAI API are state-of-the-art on challenging agentic search benchmarks.

	Research-Eval Reka	FRAMES	X Browse*
Score	Avg. Cost	Score	Avg. Cost

Grok 4.1 Fast Agent Tools API	63.9	$0.046	87.6
GPT-5	45.5	$0.107	86.0
Claude Sonnet 4.5	41.2	$0.065	85.0
Gemini 3 Pro	55.9	-	90.9

*X Browse is an internal benchmark that evaluates an agent's multihop search and browsing capabilities on X.

Grok 4.1 Fast sets a new standard in factuality, cutting the hallucination rate in half compared to Grok 4 Fast while still delivering performance on par with Grok 4 when evaluated on FActScore.