Foundry's value to xAI is also what makes Grok easy to drop there. Microsoft's model materials put Grok 4.6 on the same page as models from OpenAI, Anthropic, Meta, Mistral, DeepSeek, Cohere and Hugging Face [8], and the platform's own pitch is that customers evaluate models against their own data and pick among deployment options, with workload-specific testing and managed endpoints [12]. Any buyer who actually runs that comparison has built a swap harness rather than a relationship. Distribution bought xAI a seat in enterprise evaluations; it did not buy a default.
The detail worth reading before signing anything is the label. The Bedrock deployment xAI announced on August 19 was generally available [10]. The Foundry listing is a public preview [1]. Same weights, same published price sheet, different production footing depending on which cloud account the invoice runs through. The supplied announcements do not establish rate limits, regional availability or service-level commitments for the Foundry preview [11], so an Azure team can evaluate now and still be waiting for the numbers that determine whether the thing can carry a queue.
There is an oddity in the ownership picture. SpaceX announced its acquisition of xAI on February 2, 2026 [5], and Musk is still selling Grok through Amazon, Google and Microsoft [6]. Owning the compute underneath a model does nothing about the fact that the customer's identity controls, billing and governance already live in someone else's subscription. That is the friction Foundry removes, and it is why vertical integration upstream has not reduced the need to appear in three rival catalogs.
Then the arithmetic. API pricing starts at $2 per million input tokens and $6 per million output tokens [7], which is three to one against generated text [14]. The 500,000-token context window and the four reasoning settings, low through xhigh [3], are the cheap half of that bill. xAI says the model was trained for sustained agent work: navigating codebases, researching unfamiliar subjects, building applications and checking its own output [13], and it is pitched at coding agents, engineering copilots and enterprise automation that has to survive long sequences of steps [17]. Those workloads spend their money on the output side. A large window is close to free to fill; an agent that re-reads and verifies its own work is not.
What none of this settles is whether the model is the right one. Microsoft notes that the benchmark results in its Foundry announcement come from the model provider, and neither announcement supplies independent Azure customer results [9]. The honest read is that Foundry has removed the integration argument and left the evaluation argument entirely intact. Fourteen days from release to fourth catalog [15] [16] compresses procurement, not judgement.
