xAI 在发布一周后将 Grok 4.6 部署到 Amazon Bedrock

Why it matters

Bedrock gives Musk's xAI a direct route into AWS procurement and infrastructure. The real test is whether Grok 4.6's agentic performance earns its higher price against both Grok 4.3 and rival models already sold through Amazon.

Elon Musk (@elonmusk)'s xAI made Grok 4.6 generally available through Amazon Bedrock on Wednesday, giving AWS developers access to the model seven days after its original release.

The August 19 announcement puts Grok 4.6 inside a purchasing and deployment channel already used by enterprise engineering teams. xAI priced Bedrock access at $2 per 1 million input tokens and $6 per 1 million output tokens, matching the standard rates listed in its own launch materials. xAI says Grok 4.6 has a 500,000-token context window and four reasoning settings: low, medium, high and xhigh. (x.ai)

对于 Musk 来说，将产品列在 Bedrock 上延续了一个熟悉的剧本：在基础设施上大量投入，快速发布，然后把产品放到客户已经在使用的地方。推动这一策略的财务结构在今年早些时候发生了变化。xAI 在 1 月 6 日宣布完成 200 亿美元的 E 轮融资，而根据 SpaceX 的要约文件，SpaceX 于 2 月 2 日完成了对 xAI 的收购。这次收购将一个本已资本密集的 AI 开发者并入了 Musk 更大的工业体系。(x.ai)

A faster route into enterprise accounts

Amazon Bedrock matters because it shortens the commercial distance between a model release and an enterprise deployment. AWS customers can use an existing cloud purchasing and deployment path instead of arranging a separate model integration from scratch.

That can reduce work xAI would otherwise have to handle directly, including separate vendor onboarding and another layer of cloud integration. It also puts Grok 4.6 into a managed catalog where AWS markets the ability to evaluate and switch among models from multiple providers. (AWS)

这一安排对双方都有利。xAI 无需从头建立每一段企业关系就能进入 AWS 账户；Amazon 则为其目录增加了另一个旗舰模型，而该目录的吸引力取决于是否提供足够多可信的替代选项，让客户无需离开 Bedrock 就能进行试验。

新版本增加了 xAI 的最高推理设置，并以适用于长期运行的代理、交互式和视觉工作为卖点。Bedrock 的公告列出了 500,000-token 的上下文窗口和 2 美元/6 美元 的定价，但没有提供 Grok 4.6 的独立延迟、可靠性或使用数据。

这种对比反映了开发者面临的选择。上下文长度只是众多规格之一。构建编码代理或研究系统的团队还关心模型在长序列操作中保持任务、使用工具和从错误中恢复的能力。即便相比之前在 Bedrock 草案比较中为 Grok 4.3 列出的 1,000,000-token 容量更小，Grok 4.6 仍是 xAI 试图在这些更难的工作负载上收取更高费用的尝试，旨在 command a 更高价格。

What xAI trained Grok 4.6 to do

In the August 12 model announcement, xAI said Grok 4.6 received a longer supplemental training run than Grok 4.5. The process used model-generated reasoning data, engineering data, supervised fine-tuning trajectories regenerated by Grok 4.5, and reinforcement-learning tasks covering coding, knowledge work, web development, computer-aided design and kernel optimization.

xAI describes the result as a model for researching unfamiliar subjects, working across codebases and turning broad product ideas into functioning applications. xAI also says it observed more self-testing and verification during longer tasks. Those are xAI's findings from its own testing. The Bedrock announcement provides no separate AWS evaluation, latency measurements or production reliability data for Grok 4.6. (x.ai)

在 8 月 12 日的模型公告中，xAI 表示 Grok 4.6 比 Grok 4.5 进行了更长时间的补充训练。该过程使用了模型生成的推理数据、工程数据、由 Grok 4.5 再生的监督微调轨迹，以及覆盖编码、知识工作、Web 开发、计算机辅助设计和内核优化的强化学习任务。

xAI 将结果描述为一个用于研究不熟悉主题、跨代码库工作并将宽泛的产品想法转化为可运行应用的模型。xAI 还称在更长的任务中观察到更多的自我测试和验证。这些结论来自 xAI 自己的测试。Bedrock 的公告没有提供 AWS 独立的评估、延迟测量或 Grok 4.6 的生产可靠性数据。(x.ai)

可配置的推理等级为开发者提供了直接的成本和延迟控制。常规请求可以在低努力等级下运行，而困难的编码或研究任务则可以使用 high 或 xhigh。更高的推理努力可以提高多步工作中的表现，但也可能产生更多需计费的输出 token。在每百万输出 token 收费 6 美元的情况下，一个反复规划、调用工具、检查其工作并修正答案的代理的费用，可能远超其初始提示所显示的成本。

Bedrock 的公告表示 Grok 4.6 在受支持的 AWS 区域对所有开发者开放，并将开发者指向 AWS 文档。它并未列出模型启用的确切区域、接口或服务层级。(x.ai)

Distribution is becoming part of the model release

Grok 4.6 already had broad developer distribution before the Bedrock listing. xAI launched it through its own API, Cursor and Grok Build, with additional availability through OpenRouter, Vercel and Cloudflare. Amazon adds the channel most closely tied to established enterprise cloud budgets. (x.ai)

That speed is deliberate. Frontier model developers are compressing the gap between training a model and making it available through the platforms where developers buy inference. A model that performs well in a benchmark still has to clear procurement, security and infrastructure requirements before it can win recurring production traffic.

Musk has funded xAI to compete on both compute and reach. In its Series E announcement, xAI said it ended 2025 with more than 1 million H100-equivalent GPUs across its Colossus facilities and approximately 600 million monthly active users across X and Grok. Those figures are xAI's own measurements, and they describe potential distribution rather than paid enterprise adoption. The Bedrock launch addresses the commercial side of that equation by placing Grok 4.6 closer to the engineers and cloud accounts that can turn model usage into durable revenue. (x.ai)

在被列入 Bedrock 之前，Grok 4.6 已经在开发者中有了广泛分发。xAI 通过其自有 API、Cursor 和 Grok Build 推出该模型，并通过 OpenRouter、Vercel 和 Cloudflare 提供额外接入。Amazon 补充了与既有企业云预算最紧密相关的渠道。(x.ai)

这种速度是刻意为之。前沿模型开发者正在缩短从训练模型到通过开发者购买推理的平台上可用之间的差距。一个在基准测试中表现良好的模型，仍需通过采购、安全和基础设施要求，才能赢得经常性的生产流量。

Musk 资助 xAI 在计算能力和覆盖范围两方面竞争。xAI 在其 E 轮公告中称，到 2025 年底，其 Colossus 设施中拥有超过 100 万台 H100 等效 GPU，并且 X 和 Grok 的月活跃用户约为 6 亿。这些数字来自 xAI 自身统计，描述的是潜在分发能力，而非付费的企业采用。Bedrock 的上线通过将 Grok 4.6 更靠近能够把模型使用转化为持续收入的工程师和云账户，解决了这一方程的商业层面。(x.ai)