Key takeaways

• Grok 4.6 is a frontier Transformer-based model by xAI, designed for long-running agents, coding, knowledge work, and ambitious interactive/visual projects, building on Grok 4.5 with improved supervised fine-tuning and reinforcement learning.
• The model supports multimodal inputs (text+image+file) to text output, features configurable reasoning effort (low, medium, high, xhigh), and includes agentic tool calling (web search, X search, code execution).
• Grok 4.6 excels in mathematical problem-solving, achieving 96.5% accuracy on the MATH-500 benchmark, demonstrating strong logical deduction and detailed explanations, and proficiency in complex algebraic manipulations, geometric theorems, and number theory.
• Limitations include minor errors in algebraic simplification, occasional misinterpretation of problem constraints, and calculation errors. It struggles with interpreting visual diagrams or complex contextual information in some math problems.
• The model performs poorly on the 'Humanity's Last Exam' dataset with a very low accuracy of 2.48%, indicating significant limitations in general knowledge or complex reasoning tasks beyond specialized mathematics.

Technical specifications

Context length

500k tokens

Max output

—

Architecture

Transformer

Modalities

text+image+file->text

Parameters

~1500B

Open weights

No

Licence

Proprietary

Regions

USA

Vendor

xAI

Provider model key

x-ai/grok-4.6