SpaceXAI, the company formerly known as xAI, released its Grok 4.6 model on August 12, positioning it as a frontier system built to sustain AI agents through many steps of a task rather than short, single-turn replies, the company said in its launch announcement.
What’s new
Grok 4.6 builds on July’s Grok 4.5 release. xAI said the model went through longer supplemental training on curated data, improving its ability to research unfamiliar topics, structure applications, implement core interactions, and iterate through feedback cycles — what the company describes as turning “a broad product idea into a working first version.” It also claims better self-testing and verification during long coding sessions, and stronger first passes on visual and interactive projects.
The model takes a 500,000-token context window, accepts text and image input with text-only output, and carries a February 1, 2026 knowledge cutoff, according to xAI’s developer documentation.
Benchmarks and pricing
On the Artificial Analysis Intelligence Index, a third-party benchmark aggregator, Grok 4.6 scored 61 — up from 56 for Grok 4.5 and matching OpenAI’s GPT-5.6 Sol tier, introduced earlier this year. xAI also reported gains on agentic coding benchmarks such as GDPval-AA v2, where the model posted an Elo of 1753 versus 1526 for its predecessor.
Pricing starts at $2 per million input tokens and $6 per million output tokens, with a faster variant costing double, per xAI’s pricing page as of August 2026. Grok 4.6 is now the default model in Grok Build, the company’s own coding tool, and ships in the Cursor code editor “on all plans.” It’s also reachable through the xAI API and routable via OpenRouter, Vercel, and Cloudflare. xAI is offering double the usual included usage in Grok Build and Cursor for the model’s first week.
The release follows SpaceXAI’s July launch of Grok Bot, a set of always-on agents for macOS and iOS, underscoring the company’s push toward tools that keep working on tasks in the background rather than waiting on a user’s next prompt.