xAI launches Grok 4.6 and Grok Bot, positioning the model as a leading coding and agentic knowledge work system

xAI introduced Grok 4.6, a 1.5T-parameter model designed to enhance long-horizon agentic tasks and knowledge work. The model builds on Grok 4.5 with a focus on interactive and visual work, supported by a longer supplemental training run and curated model-generated data for reasoning and technical concepts.

Grok 4.6 underwent additional training with high-quality engineering data and an improved optimizer, followed by supervised fine-tuning (SFT) and reinforcement learning (RL) stages. The training process included regenerating SFT trajectories across domains such as STEM, software engineering, and knowledge work, with model-based checks to filter problematic traces.

The model is trained on agentic RL tasks spanning knowledge work, general coding, kernel optimization, web development, and computer-aided design. xAI claims Grok 4.6 delivers strong performance while maintaining efficiency, positioning it as a top contender in agentic knowledge work.

xAI also launched Grok Bot, an AI teammate available in early beta. Grok Bot integrates with user tools, performs autonomous work, and returns completed tasks, marking a new entrant in the AI teammate category.

Independent evaluations from Artificial Analysis place Grok 4.6 at 61 on the Intelligence Index, roughly in line with GPT-5.6 Sol Max and behind Claude Opus/Fable. However, the model excels in agentic benchmarks, achieving 88.4% on Terminal-Bench v2.1 and 1753 GDPval-AA v2 Elo, with competitive performance on the AA-Briefcase agentic knowledge work benchmark.

Grok 4.6 is priced at $2 per 1M input tokens and $6 per 1M output tokens, significantly undercutting frontier peers. Practitioners have already adopted it as a default for coding and bug-finding workloads, according to early user reports.

xAI indicated that Grok 4.7 is in development, with initial training complete and supplemental training planned on SpaceX internal data. The company also emphasized the model's self-testing behavior during long tasks, which enhances reliability in autonomous workflows.