xAI 发布了 Grok 4.6 大语言模型 [1]，这是对 Grok 4.5 的升级版本 [1]。该模型在 Artificial Analysis Intelligence Index 上获得 61 分 [2]，与 OpenAI 的 GPT-5.6 Sol 相当 [2]，仅次于 Anthropic 的 Claude Opus 5 [2]。相比 Grok 4.5，该版本在该指数上提升了 5 分，相比 Grok 4.3 提升了 23 分 [2]。
在功能定位上，Grok 4.6 重点关注长时间运行的智能代理和交互视觉工作 [1]，在多个代码编写和知识工作基准上达到前沿水平 [1]。其在 τ³-Banking 评分达 50.7%，Terminal-Bench v2.1 评分达 88.4% [2]，均处于领先水平。该模型拥有 500k tokens 的上下文窗口 [2]，在 AA-Briefcase 基准上的平均用时约 53 轮和 0.5B 输入 tokens，相比 Claude Opus 5 的 103 轮和 2.0B 输入 tokens 显著更高效 [2]。
定价方面，Grok 4.6 保持每百万输入 token 2 美元、每百万输出 token 6 美元的价格 [1][2]，快速版本价格翻倍 [1]，缓存命中折扣调整为每百万 token 0.5 美元 [2]。这使其成本效益每项任务仅为 0.84 美元，相比 Claude Opus 5（$5/$25）和 GPT-5.6 Sol（$5/$30）降低超过 60% [2]。该模型已在 Cursor 和 Grok Build 平台上线 [1]，首周用户可获得 2 倍的免费使用量 [1]，同时可通过 API 和 OpenRouter、Vercel、Cloudflare 等合作伙伴获得 [1]。
xAI has released Grok 4.6, an upgraded language model that achieves frontier performance on multiple coding and knowledge work benchmarks [1]. The model scores 61 on the Artificial Analysis Intelligence Index, matching OpenAI's GPT-5.6 Sol and ranking second only to Anthropic's Claude Opus 5 [2]. Grok 4.6 emphasizes long-running intelligent agents and interactive visual tasks, with particular strength in agent performance [1][2].
The model is priced at $2 per million input tokens and $6 per million output tokens [1][2]. A faster variant costs twice as much [1]. During the first week of availability on Grok Build and Cursor, users receive double the free usage allocation [1]. The model features a 500,000-token context window and demonstrates significant efficiency gains on extended knowledge work tasks compared to competing models [2]. On the AA-Briefcase benchmark, Grok 4.6 achieves an Elo score of 1577, completing tasks in approximately 53 rounds and consuming roughly 0.5 billion input tokens, compared to Claude Opus 5's approximately 103 rounds and 2.0 billion tokens [2].
Grok 4.6 is available through API access and via partnerships with OpenRouter, Vercel, and Cloudflare [1]. The model's cost-effectiveness stands at $0.84 per task, representing over 60 percent savings compared to Claude Opus 5 ($5/$25) and GPT-5.6 Sol ($5/$30) [2]. Cache hit discounts are set at $0.50 per million tokens [2].