Grok 4.6 Hits Microsoft Foundry Preview: What Ships

Grok 4.6 is now in public preview in Microsoft Foundry as an Azure Direct Model, with a 200K token context window and per-call reasoning depth.

One source dates the model's release to August 12, 2026, attributing it to xAI, alongside Meta's Muse Spark 1.2 on August 5 — two frontier releases in the same month, both aimed at autonomous coding rather than conversation.

What ships in the preview

A .NET developer walkthrough published this week lists the specifics worth knowing before you deploy it:

• Selectable reasoning effort per call — low, medium, high, or xhigh, defaulting to high. The point, per the write-up, is not paying maximum-reasoning cost on every request regardless of whether the task needs it.
• 200K token context window at launch. The author calls this "solid for most agentic and document-analysis workloads" and advises setting expectations up front if your scenario needs more.
• Multimodal input — text and images, so document-heavy and screenshot-heavy workflows do not need a separate vision pipeline bolted on.
• Still preview. The write-up is explicit: validate against your own prompts, tools, and safety thresholds before anything production-sensitive touches it.

On pricing, the post describes Grok 4.6 as "positioned as the value-tier frontier option — frontier-class reasoning at a materially lower cost per task than comparable models." That is positioning language, not a published rate card; no per-token figures appear in the source, so treat the cost claim as unverified until Microsoft publishes Foundry pricing for the model.

The deployment wrinkle

The one integration detail that will cost you time: Grok is a partner (MaaS) model, not a native Azure OpenAI deployment, so it cannot be reached through AzureOpenAIClient's /openai/deployments/... path. The walkthrough routes around it with the generic OpenAI ChatClient pointed at the Foundry model endpoint, using a bearer-token policy for Entra ID auth.

Deployment itself is ordinary — grok-4.6 comes from the Foundry Model Catalog, currently Global Standard deployment only. And once it is up, the author notes it "slots into Foundry the same way every other model does — same IChatClient abstraction, same deployment pattern," which makes adding it to an existing evaluation pipeline a config change rather than a rewrite.

Is it worth trying?

If you already run models through Foundry, the cost of evaluating this one is close to zero — a deployment and a config entry. That is the strongest argument for looking now rather than waiting for general availability.

If you are choosing a model for long-running agent work specifically, the honest answer is that the public material is a vendor-shaped feature list, not a benchmark. The claims about long-horizon execution, tool reliability, and error recovery come from the launch positioning; no independent numbers accompany them in the source. Selectable reasoning effort is the one feature with a clear, checkable payoff, because you can measure the cost difference between low and xhigh on your own workload in an afternoon.

The thing to watch is whether independent coding benchmarks pick Grok 4.6 up now that it is reachable through a mainstream enterprise endpoint. Foundry availability puts it in front of teams that would never have provisioned it directly — which is usually when the real numbers start appearing.