Greater continuity across complex tasks with Grok 4.6

Grok 4.6 improves long-running agentic tasks, software development, coding, and interactive projects, with gains across several benchmarks.

Grok 4.6 builds on Grok 4.5 with a stronger focus on long-running agentic tasks, software development, and interactive or visual projects. The model is designed to stay engaged across workflows involving many steps, whether researching a topic, analyzing information, working across a codebase, or progressively building an application.

xAI says Grok 4.6 underwent a longer supplemental training phase than Grok 4.5. The training mix includes model-generated data for reasoning and advanced technical concepts, engineering data, followed by SFT and reinforcement learning stages. Grok 4.5 was also used to regenerate certain SFT trajectories across STEM, software engineering, and knowledge work, before problematic traces were filtered out.

The reinforcement learning stage covers several agentic environments, including general coding, kernel optimization, web development, computer-aided design, and various knowledge-work tasks. On the long-running tasks tested by xAI, Grok 4.6 shows more self-testing and verification behavior before moving forward. The company also reports stronger first passes on visual and interactive projects, with the model able to build an initial working version and refine it through multiple rounds of feedback.

Published results place Grok 4.6 High at 61 on the Artificial Analysis Intelligence Index, matching GPT-5.6 Sol Max and sitting just behind Fable 5 Max at 62. The model also scores 1,753 on GDPVal-AA v2, 69.9% on CursorBench v3.2, 65.9% on DeepSWE v1.1, and 61.3% on FrontierCode v1.1 Extended. It shows clear gains over Grok 4.5 High across these evaluations, although it does not consistently take the top position against the other models compared.

This focus is particularly visible in agentic-work benchmarks. Grok 4.6 reaches 57.5% on APEX-Agents, 56.4% on APEX-SWE, and 26% on Terminal-Bench v3.0. On AA-Briefcase, which focuses on professional tasks, xAI reports a score of 1,577, slightly ahead of Fable 5 Max at 1,574 and GPT-5.6 Sol Max at 1,502 in the published table.

The model is available in Cursor and Grok Build, as well as through the API and partners including OpenRouter, Vercel, and Cloudflare. Pricing starts at $2 per million input tokens and $6 per million output tokens. A Fast variant is also available at twice the price.