Overview
<cite index="16-1,16-2">Grok 4.6, xAI's frontier model for coding, agentic tasks, and knowledge work, became generally available on the xAI application programming interface (API) on August 12, 2026, offering a 500,000-token context window, text and image inputs, and text-only output.</cite> The release is a post-training upgrade rather than a new base model: <cite index="14-1">according to the official x.ai writeup, Grok 4.6 ran a longer supplemental training pass than its predecessor, using curated model-generated data, high-quality engineering data, and an improved optimizer.</cite>
Benchmark Performance
<cite index="18-7">On xAI's launch table, Grok 4.6 (High) scores 61 on the Artificial Analysis Intelligence Index, up from 56 for Grok 4.5 and tied with GPT-5.6 Sol Max.</cite> <cite index="15-4">On that index, the model sits level with GPT-5.6 Sol and one point behind Claude Fable 5, placing it among the top handful of frontier models.</cite> Specific task benchmarks show a mixed picture: <cite index="2-5">Grok 4.6 leads on GDPval-AA v2 (1,753 Elo, versus 1,526 for Grok 4.5) and AA-Briefcase (1,577, versus 1,313).</cite> On coding evaluations, <cite index="2-7">DeepSWE v1.1 lands at 65.9%, up 11.9 points generationally but behind GPT-5.6 Sol Max at 73%.</cite> <cite index="14-4">No independent third-party replication of the benchmark numbers existed as of August 13, 2026; the figures are published by xAI or repeated by commentators.</cite>
Pricing Structure
<cite index="6-8">For prompts below 200,000 input tokens, standard API pricing starts at $2 per million input tokens, $0.50 per million cached input tokens, and $6 per million output tokens.</cite> <cite index="1-15">In xAI's long-context band at 200,000 tokens, those rates double to $4, $1, and $12 per million tokens respectively, and the higher rates apply to all tokens in the request.</cite> <cite index="8-9">Standard token rates remain level with Grok 4.5, while cached input becomes more expensive and the 200,000-token long-context threshold can double the entire request's rate.</cite>
Model Specifications
<cite index="18-4,18-5">The model carries a February 1, 2026 knowledge cutoff; reasoning effort now supports low, medium, high (default), and a new "xhigh" level.</cite> <cite index="18-6">xAI did not publish a parameter count for Grok 4.6.</cite> On the training side, <cite index="14-11">NVIDIA stated that Grok 4.6 was trained and runs on NVIDIA GB300 NVL72 with NVLink.</cite>
Distribution and Availability
<cite index="15-5">Grok 4.6 is available through Cursor, the coding platform that xAI acquired for $60 billion in June 2026, through xAI's own Grok Build programming tool, and through the xAI API.</cite> <cite index="14-2">It is also available on OpenRouter, Vercel, and Cloudflare, per official announcements.</cite> <cite index="2-13">Grok Build and Cursor are offering 2× included usage for the first week of availability.</cite>
Context Window Caveat
<cite index="1-10">The 500,000-token context window did not expand from Grok 4.5, which was already listed at 500K.</cite> The pricing design, however, introduces a meaningful cost discontinuity: <cite index="4-1">pricing doubles once a request crosses 200,000 tokens, a threshold that represents the detail most likely to surprise developers at billing time.</cite> xAI's own documentation recommends use of prompt caching and context compaction to manage costs in long-running agentic workflows.