Grok 4.6

What it is

Grok 4.6 is a large language model released by xAI on 12 August 2026. It builds on Grok 4.5, with the company emphasising long-running agents and interactive and visual work as the design targets.^[1] The model has a context window of 500,000 tokens, accepts text and images as input, and returns text.^[2]

How it was trained

According to xAI, Grok 4.6 received a longer supplemental training run than Grok 4.5. That run used curated model-generated data aimed at reasoning and advanced technical concepts, high-quality engineering data, and an improved optimiser and training recipe.^[1]

Part of the pipeline reused the previous model. Grok 4.5 was used to regenerate the supervised fine-tuning trajectories across reasoning efforts, agent harnesses and domains such as STEM, software engineering and knowledge work, and traces judged problematic were filtered out by model-based checks.^[1] The agentic reinforcement learning stage covered knowledge work and general coding, as well as domain-specific environments for tasks such as kernel optimisation, web development and computer-aided design.^[1]

Benchmarks

xAI reported the following results at high reasoning effort: a score of 61 on the Artificial Analysis Intelligence Index, against 56 for Grok 4.5; 69.9% on CursorBench v3.2; 65.9% on DeepSWE v1.1; 57.5% on APEX-Agents; and 26% on Terminal-Bench v3.0.^[1] The company also said that Grok 4.6 matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index, a composite of nine benchmarks.^[1] These are figures published by the developer rather than the outcome of an independent evaluation.

Price and access

The xAI API lists $2 per million input tokens, $0.50 per million cached tokens and $6 per million output tokens for prompts below 200,000 tokens; above that threshold the rates are $4, $1 and $12 respectively.^[2] Reasoning effort is a setting that controls how much internal work the model does before answering, and Grok 4.6 added an "xhigh" level above low, medium and high, with high as the default.^[2] A Fast variant is priced at twice the standard rate.^[1]

The model is available in Cursor, Grok Build, the xAI API, OpenRouter, Vercel and Cloudflare,^[1] and also through GitHub Copilot and Amazon Bedrock.^[3] For its first week it came with 2x included usage inside Grok Build and Cursor.^[1]

Place in the line

Grok 4.6 sat between Grok 4.5 and Grok 4.7, which arrived on 21 September 2026 with the same context window and the same prices.^[3][2] The recipe visible in the release notes — a further training run on top of an existing model, plus a fine-tuning set regenerated by that model — describes how this generation of the Grok line was assembled.