Foundry Model Router Expands from Two Regions to 28, Refreshing Its

What happened

Foundry Model Router Expands from Two Regions to 28, Refreshing Its Model Pool

Microsoft has expanded the Foundry model router from two regions to 28 regions for global standard deployments and 21 regions for data zone deployments. Alongside the geographic expansion, the router's model pool has been refreshed: Claude Opus 4.8 and GPT-5.6 have been added, and four deprecated models have been removed.

This article walks through the specifics of what shipped, what the changes mean for teams building on Foundry today, and what a builder needs to verify before their next deployment. It is aimed at developers, platform engineers, and architects who have Foundry model router configurations in production or are evaluating it for new projects.

How it works

Microsoft pushed two parallel changes to the Foundry model router simultaneously: a dramatic geographic expansion and a pool refresh. On the infrastructure side, router availability jumped from two regions to 28 for global standard deployments and to 21 for data zone deployments. The global standard tier covers the broader set of Azure regions and is suited for applications that do not have strict data residency requirements. The data zone tier is the smaller set, intended for workloads that must keep data within a defined geographic boundary. Both tiers grew substantially from a two-region baseline.

On the model side, Claude Opus 4.8 and GPT-5.6 entered the pool while four previously available models were removed. The four removed models are described as deprecated, which means they had already been flagged as end-of-life before this update. The additions of Claude Opus 4.8 and GPT-5.6 represent the highest-capability options now available in the router's curated selection, and their inclusion signals that the router is being kept current with the latest model generations from both Anthropic and OpenAI rather than lagging behind direct endpoint availability.

The router's behavior after this update depends entirely on how a deployment was configured before the change landed. Deployments that use the default pool, meaning no explicit subset was specified at configuration time, receive the new pool composition automatically. That means Claude Opus 4.8 and GPT-5.6 are now live and eligible for routing in those deployments without any action required from the builder. The four deprecated models are also gone from those deployments immediately.

Why it matters

Deployments that were configured with an explicit subset of models work differently. For those, the router excludes any model not already named in the subset configuration. Claude Opus 4.8 and GPT-5.6 will not be routed to until the builder explicitly adds them to the configured list. This is a deliberate design choice that prioritizes stability over automatic adoption in controlled environments, but it does mean teams expecting access to the new models must take a manual step to update their configuration.

There is no SDK or package installation step required to take advantage of the expanded region coverage or the new models. The changes are infrastructure-level and take effect on the Foundry service side. For teams on default pool deployments, the changes are already active. The new regions are available without any configuration change, and routing traffic to a newly supported region is a matter of selecting that region when creating or updating a deployment.

Who is affected

For teams on explicitly configured subsets, the upgrade path is to open the deployment configuration and add Claude Opus 4.8 and GPT-5.6 to the model list. After saving, the router will begin routing eligible requests to those models according to its standard selection logic. Teams should also audit their configurations to confirm the four deprecated models are no longer referenced in fallback logic, application-level parsing, or downstream tooling that may have been built around specific model identifiers.

The context window behavior is worth understanding carefully before relying on the new models in production. The effective context window for any router-based deployment equals the smallest context window among the models in the active pool. If a configured subset mixes Claude Opus 4.8 or GPT-5.6 with models that have smaller context limits, the router constrains the effective window to that lower bound. Teams that upgraded their context expectations based on the new models' standalone limits may find requests truncated or rejected if other models in their pool set a lower ceiling.

Removing the four deprecated models also affects any logging, monitoring, or alerting that references model identifiers by name. Pipelines that parse response metadata for model names used in cost attribution, latency tracking, or audit trails may encounter null values or unexpected identifiers. Validating that observability tooling handles the removed model identifiers gracefully before traffic is routed in production will prevent gaps in instrumentation.

What to watch next

The jump from two regions to 28 global standard regions in a single update is a large step, and the natural question is whether data zone coverage follows a similar growth curve. The data zone tier expanded to 21 regions, which is substantial but still below the global standard count. Builders with data residency constraints will want to track whether the remaining gap closes in subsequent updates or whether certain jurisdictions require separate regulatory work before they can be included.

The addition of Claude Opus 4.8 and GPT-5.6 also raises the question of cadence. If the router is now tracking major model releases from both Anthropic and OpenAI with relatively low lag, teams should expect the pool to continue turning over and should build configuration management practices that account for regular additions and removals rather than treating the pool as stable.