SpaceX AI launches Grok Bot agent; Grok 4.6 challenges frontier models on cost

SpaceX's xAI launched Grok Bot, an AI agent now in early beta that autonomously completes work within your existing tools. The agent signs in just as a human colleague would, uses those tools directly, and returns finished work—a model aligned with how practitioners have begun to think about AI teammates rather than pure text interfaces. Grok Bot runs on Grok 4.6, xAI's newest model released on August 11–12, 2026, which represents a major step up from Grok 4.5 at the same price point.

Grok 4.6 is a 1.5T parameter model built through a longer supplemental training run than its predecessor, with curated model-generated data for reasoning and advanced technical concepts, high-quality engineering data, and an improved optimizer and training recipe. The team regenerated SFT (supervised fine-tuning) trajectories across reasoning, agent harnesses, and domains including STEM, software engineering, and knowledge work, filtering out problematic traces with model-based checks. Crucially, Grok 4.6 underwent agentic reinforcement learning over a wide range of tasks: knowledge work, general coding, kernel optimization, web development, and computer-aided design.

Intelligence Index evaluation placed Grok 4.6 at a score of 61, roughly in line with GPT-5.6 Sol Max and trailing Claude Opus/Fable, but its agentic benchmarks prove decisive. Terminal-Bench v2.1 showed 88.4% performance, and GDPval-AA v2 Elo reached 1753; early Code Arena data from webdev tasks also slotted it near GPT-5.6 Sol and Claude Fable. Pricing emerged as the central competitive advantage: xAI announced $2 per 1M input tokens and $6 per 1M output tokens, materially below frontier peers. Practitioners immediately framed it as the new default for coding and bug-finding workloads, with Cognition's Pawel Huryn noting Devin's Grok availability. Early reports also highlighted increased self-testing behavior during long tasks.

Elon disclosed that Grok 4.7 is already in flight with initial training complete and supplemental training on SpaceX internal data planned. This roadmap, combined with the week's parallel releases—DeepSeek V4 Pro GA (priced around $0.435/M input and $0.87/M output with a reported 15.8% Terminal Bench increase over preview), Alibaba's Qwen3.8-Max open-weights (2.4T total / 95B active MoE, day-0 support from vLLM and vendor-specific 4-bit checkpoints), and Microsoft's MAI-Thinking-1 reasoning model—signals that cost efficiency and agentic capability have become table stakes for frontier competition.