[AINews] SpaceXAI Grok 4.6 and Grok @Bot

xAI has pushed Grok 4.6 into the center of the week’s AI chatter, and the pitch is blunt: this is the model for agentic work, not just chat. The company describes it as a 1.5T model tuned for long-running agents and more ambitious interactive and visual tasks, with an emphasis on coding, knowledge work, web development, CAD, and kernel optimization. That’s a wide brief, but the timing makes sense. The fight now is less about raw chatbot polish and more about which model can actually sit inside a workflow and keep going.

The training story points in the same direction. xAI says Grok 4.6 went through a longer supplemental run than Grok 4.5, using curated model-generated data for reasoning and technical concepts, plus engineering data, a new optimizer, and a revised training recipe. Then it regenerated SFT trajectories from Grok 4.5 across reasoning, agent harnesses, STEM, software engineering, and knowledge work, filtered out bad traces, and built an SFT checkpoint from there. In other words: more work on behavior, less on hype. The company also says Grok 4.6 is trained on agentic RL tasks spanning knowledge work, coding, and specialized environments.

Independent evaluation data is what makes this release feel more than just another model drop. Artificial Analysis put Grok 4.6 at 61 on its Intelligence Index, around GPT-5.6 Sol Max and behind Claude Opus/Fable, while also calling out strong agentic results. The numbers cited include 88.4% on Terminal-Bench v2.1, 1753 GDPval-AA v2 Elo, and competitive AA-Briefcase performance at a lower cost. Pricing matters here too: AA lists Grok 4.6 at $2 input and $6 output per 1M tokens, which is a big part of why practitioners are treating it as a default for coding and bug-finding.

That efficiency angle is probably the real story. If a model lands near the top tier while costing less, teams stop talking about benchmarks and start wiring it into actual systems. Elon also said Grok 4.7 is already in flight, with initial training complete and supplemental training on SpaceX internal data planned. So yes, the agent race just got a new front-runner, and it is being sold like infrastructure, not a demo.