Grok 4.6 arrives tuned for agents that keep working after the demo ends
Wednesday's release is a post-training upgrade, not a bigger model, and its pitch is stamina: staying with codebases, research tasks and builds across many steps. The coding benchmarks tell a more mixed story.
SpaceXAI released Grok 4.6 on Wednesday, and the framing was unambiguous: this is a model for agents that run long. The company describes a system that stays engaged with complex work across many steps, researching unfamiliar topics, moving through large codebases over multiple revisions, accepting iterative feedback, and checking its own progress before continuing, Unite.AI reported.
Under the hood this is a post-training upgrade to Grok 4.5 rather than a larger base model, built on extended training runs and reinforcement learning across software engineering, web development, and CAD, MarkTechPost reported. The context window is 500,000 tokens, input covers text and images, and a new xhigh reasoning-effort setting joins the existing low, medium, and high options, giving fleet operators another dial for trading depth against cost on a per-task basis.
Pricing is built for the same audience. Below 200,000 prompt tokens the model runs $2 per million input tokens, 50 cents cached, and $6 per million output; past that threshold the rates double, and a faster variant costs twice the standard price, per MarkTechPost. The structure rewards agents that manage their own context hygiene, and penalizes the ones that drag a whole repository into every call.
The scorecard is genuinely mixed. Grok 4.6 posts a 61 on the Artificial Analysis Intelligence Index, up five points from Grok 4.5 and tied with GPT-5.6 Sol Max, while trailing the top of that composite table, according to Unite.AI. On agentic coding the picture inverts: MarkTechPost notes the model trails on the rows engineering teams watch most, scoring 65.9 percent on DeepSWE v1.1, a large jump from 4.5 but short of rival frontier models, and 26 percent on Terminal-Bench v3.0. SpaceXAI's own emphasis on knowledge work, visual builds, and endurance over raw software-engineering scores reads as an honest map of where the model wins.
Distribution is where the release gets aggressive. Grok 4.6 went live simultaneously in the xAI API, in Cursor on all plans, and in the Grok Build app-generation tool, with routing available through OpenRouter, Vercel, and Cloudflare, and double usage credits in Cursor and Grok Build during launch week. There are no open weights and no self-hosting path.
The strategy is legible: if capability tables stay tight, win on price, placement, and staying power. With Cursor now in-house and Grok Bot fleets rolling out on dedicated cloud machines, Grok 4.6 is less a standalone launch than the engine SpaceXAI intends to run underneath a full agent-as-a-service stack, sold by the seat above and by the token below.