xAI 发布 Grok 4.6

2026 年 8 月 12 日，xAI 发布 Grok 4.6。总参数约 1.5T 的 MoE 旗舰，上下文 500K，支持文本与图像输入——这些数字放在 7 月的 Grok 4.5 之后并不算意外，真正值得注意的是发布当天的两件事：一是马斯克宣称它「客观排名第一」，二是连续第二次没有附带官方 model card。

先看能力。官方口径里，Grok 4.6 在 Artificial Analysis 综合智能指数上与 GPT-5.6 Sol 持平（61 分），DeepSWE 65.9%、Terminal-Bench 26.0%、APEX-Agents 57.5%。这些数字需要带着厂商自述的边界读——在纯编程任务上，它仍落后于 GPT-5.6 Sol Max 与 Claude Fable 5 Max。xAI 的差异化不在「最强」，而在「够强且便宜」：API 定价每百万输入 2 美元、输出 6 美元（200K 以下），大约是 GPT-5.6 Sol（5/15）与 Claude Opus 5（5/25）的一半。对跑长程智能体任务的团队来说，这个价格差会直接写进账单。

更值得读的是发布渠道。Grok 4.6 同步上线 Cursor 与 Grok Build，首周双倍内含用量。6 月 xAI 收购 Cursor 时，很多人还在猜这笔交易的意义；一个月后答案浮出水面——xAI 不再满足于「X 上的聊天模型」，而是要把编程入口、智能体工作流和旗舰模型绑成一条产品线。Grok Build 是自家入口，Cursor 是收购来的入口，API 是开放入口，三路并进。

但同一天的另一面同样重要：连续第二次没有 model card。Forkast 等媒体直接质疑：对搭建自主智能体的企业来说，不知道模型在什么条件下会做什么、不会做什么，本身就是风险。能力先行、文档缺席，正在成为 xAI 发布节奏的固定特征——这与其说是疏忽，不如说是选择：用速度换透明，用价格换份额。

Grok 4.6 的意义不在单个数字。它把 xAI 的旗舰从「聊天 + 搜索」推进到「长程智能体 + 编程入口」，并用低于竞品一半以上的价格重新定义旗舰定价。对开发者，多了一个价格友好的旗舰选项；对行业，「能力溢价」的定价逻辑第一次被正面挑战。至于「客观排名第一」的说法，留给榜单去裁决——但「便宜一半的旗舰」这件事，不需要任何榜单来证明。

On August 12, 2026, xAI released Grok 4.6. A ~1.5T-parameter MoE flagship with a 500K context and text-plus-image input—these numbers were not surprising after Grok 4.5 in July. What stood out was two things on launch day: Musk's claim that it is "objectively number one," and the second consecutive release without an official model card.

On capability: by xAI's own account, Grok 4.6 matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index (61), with DeepSWE 65.9%, Terminal-Bench 26.0%, and APEX-Agents 57.5%. These numbers need the vendor-self-reported caveat—on pure coding tasks it still trails GPT-5.6 Sol Max and Claude Fable 5 Max. xAI's differentiation is not "strongest" but "strong enough and cheap": API pricing at $2 per million input and $6 per million output (under 200K), roughly half of GPT-5.6 Sol ($5/$15) and Claude Opus 5 ($5/$25). For teams running long-horizon agent tasks, that gap lands directly on the invoice.

The distribution channels deserve equal attention. Grok 4.6 launched on Cursor and Grok Build simultaneously, with double included usage in the first week. When xAI acquired Cursor in June, many wondered what the deal meant; a month later the answer surfaced—xAI no longer wants to be "the chat model on X." It is binding the coding entry point, agent workflows, and the flagship into one product line: Grok Build as the home door, Cursor as the acquired door, the API as the open door.

But the other face of the same day matters too: no model card, again. Forkast and others asked directly: for enterprises building autonomous agents, not knowing what a model will and won't do under what conditions is itself a risk. Capability first, documentation absent is becoming a fixed feature of xAI's release cadence—less an oversight than a choice: speed over transparency, price over share.

Grok 4.6's meaning is not in any single number. It moved xAI's flagship from chat-plus-search into long-horizon agents and a coding entry point, and redefined flagship pricing at less than half of rivals. For developers, there is a new price-friendly flagship option; for the industry, the "capability premium" pricing logic was challenged head-on for the first time. As for "objectively number one"—leave that to the leaderboards. "A flagship at half price" needs no leaderboard to prove.