Microsoft Foundry September 2026: GPT-6 GA, Agent Routines and Foundry Local 2.1

Alongside cloud model expansions, the update delivers Insights in public preview for trace-based debugging, adds session affinity and routing telemetry to Model Router, and ships Foundry Local 2.1.0 with bring-your-own-model (BYOM) support for Windows and cross-platform developers working with ONNX Runtime GenAI. Here is a complete technical breakdown of what shipped, where the architectural boundaries sit, and what you need to verify before upgrading your SDKs or production pipelines.

The GPT-6 Lineup Reaches General Availability​

The headline model news for September is the general availability of three offerings in the GPT-6 family: GPT-6 Astra, GPT-6 Luna, and the late-month arrival GPT-6.1 Sol. Rather than forcing every workload through a single frontier bottleneck, the lineup splits cleanly across capability and token economics.

All three models feature a 1.05-million-token context window and support up to 128,000 output tokens (where output and reasoning tokens share a single budget). Across the board, they support the Responses API, Chat Completions, streaming, function calling, structured outputs, and image inputs with text outputs. However, developers need to watch one critical API boundary: tool calls on GPT-6.1 Sol require the Responses API.

Note: Pricing reflects September-announced USD rates per 1 million tokens for Global Standard short-context requests. Long-context requests carry higher rates, and regional or Data Zone deployments vary.

According to Microsoft’s launch notes, GPT-6.1 Sol delivers stronger coding, computer-use, and professional task performance than its predecessor GPT-6 Sol while keeping uncached input and output prices identical and cutting cached-input pricing by 50%. Because that comparison comes directly from the release announcement rather than independent third-party testing, engineering teams should benchmark GPT-6.1 Sol against their own repositories and tool-calling prompts before migrating production traffic.

Claude Opus 5.5 and MAI Model Updates​

Beyond the OpenAI lineup, September expanded third-party and Microsoft AI (MAI) model options inside Foundry:

• Claude Opus 5.5 (Announced September 24): Available in Foundry via either Azure-hosted or Anthropic-hosted deployment paths. Developers must verify which hosting option and API contract their environment uses—do not assume standard OpenAI client parameters transfer unchanged.
• MAI-Image-2.6 and MAI-Image-2.6-Flash (September 4): Entered public preview for developer availability in Foundry, supporting tasks like multi-reference editing with synthetic product imagery.
• MAI-Transcribe-2 (September 3): Confirmed for Foundry demo availability, allowing teams to evaluate speaker diarization, word-level timestamps, and domain terminology against fixed audio clips. Note that demo availability does not represent a universal API or deployment GA transition.

Model Router: Session Affinity, Metadata, and 32 Regions​

When an application relies on Model Router to dynamically pick the right model for a prompt, multi-turn conversations can get awkward if the router jumps between different underlying models on every turn. In September, Microsoft added two preview controls for Chat Completions on Model Router:

1. Session Affinity (Preview): Applications can pass an opaque session ID across related Chat Completions turns to ask Model Router to prefer the same eligible model. While this can improve prompt-cache reuse, Microsoft Learn’s Model Router documentation notes that neither a cache hit nor the same model is guaranteed—safety filters, capability requirements, quota limits, availability, and automatic failover still take precedence.
2. Per-Request Routing Metadata (Preview): When clients opt into the preview response contract, responses identify the serving model and return observability metadata including routing mode, routing-trace latency, ordered model attempts, HTTP status codes, and affinity decisions (whether a session initialized, retained, or switched models). Because preview fields are optional, client code should parse them defensively.

Model Router also expanded from 28 to 32 Azure regions in September, adding Canada Central, North Europe, Norway East, and UAE North. North Europe and Norway East support both Global Standard and Data Zone Standard deployments, while Canada Central and UAE North support Global Standard only.

Section Summary: The GPT-6 family (Astra, Sol, Luna) is now GA with a 1.05M context window and tiered pricing from $0.10 to $10.00 per million input tokens, while Model Router adds preview session affinity and trace metadata across 32 Azure regions.

Agents and Foundry Agent Service: Routines, Tool Search, and Insights​

Building an agent that works in a local notebook is easy; running one unattended every Monday morning without blowing up your token budget or spamming a Microsoft Teams channel is an architecture problem. On September 24, Microsoft pushed several foundational pieces of the Foundry Agent Service into General Availability and public preview.

Routines and Tool Search Hit GA​

• Routines (GA): A Routine starts a single agent from a single trigger—a timer, a recurring schedule, a GitHub issue event, or a supported Microsoft Teams channel event. Its configuration, identity, and run history remain attached to the Foundry project. Importantly, Microsoft draws a strict architectural boundary here: Routines replace external cron schedulers for starting a job, but they do not replace multi-agent orchestration or human approval workflows. Additionally, the agent-scheduled reminder tool remains in preview.
• Tool Search (GA): If an agent has access to dozens of tools, stuffing every JSON schema into every prompt wastes context and drives up input-token billing. Toolboxes now expose on-demand discovery and invocation through tool_search and call_tool. While Hosted-Agent Toolboxes were already GA, Prompt-Agent Toolboxes entered public preview in September.
• Insights (Public Preview): Once an agent is running, Insights analyzes saved traces to group recurring behavioral problems into reviewable findings, complete with supporting trace evidence, likely causes, and proposed fixes. For example, in Microsoft’s internal changelog workflow, Insights flagged sample-heavy prompts sent after a batch was already ruled out, large prompts retried immediately after rate-limit errors, and an earlier unconfirmed delivery receipt. Insights is designed for a human-controlled improvement loop—engineers should inspect the linked trace and test the proposed change against fixed inputs rather than applying changes automatically.
• Voice Agents, Long-Running Resilience, and A2A: Agent-to-Agent (A2A) delegation is now GA alongside Tool Search. Meanwhile, Voice agents in Foundry Agent Service (distinct from the Hosted Agents with Voice Live integration covered in summer releases) and long-running hosted-agent resilience entered public preview. Long-running resilience requires explicit background-response and server configuration, and developers must still implement application checkpointing and idempotency so reconnecting a dropped client does not re-trigger external side effects.
• Entra and Lifecycle Targets: Availability for Entra and Agent 365 lifecycle operations was confirmed by September 24 (without a separate maturity label), allowing admins to verify that lifecycle changes affect an agent's runtime and not just its directory record. Conversely, rubric evaluation, synthetic/trace-derived dataset generation, and Agent Optimizer GA were only announced as targets for "later this month"—even though related GA API surfaces appeared in the Python Projects SDK 2.8.0 on October 2, they remain September targets rather than confirmed September service releases. The announcement also introduced the open-source run-assert-eval skill for connecting risk discovery and policy-based evaluation.

Real-World Architecture: The Two-Agent Developer Changelog​

To demonstrate how these pieces fit together without relying on hand-wavy magic, the Microsoft Foundry Blog detailed a two-agent developer changelog workflow tested through an October 13 demonstration:

1. The Producer (depend-agent): Built with the GitHub Copilot SDK, triggered by a Foundry Routine, and equipped with Tool Search over seven scoped, read-only GitHub operations. Application code—not model judgment—restricts the agent to approved repositories, file paths, and PR-read methods.
2. The Receiver (maf-update-agent): Built with the Microsoft Agent Framework, which retrieves the verified report from the producer via the Invocations protocol and delivers a single Adaptive Card to a fixed, code-approved Microsoft Teams destination via the native Activity protocol.

Why split the protocols? Microsoft’s breakdown highlights a vital design rule for enterprise developers: choose the external protocol for the caller’s contract, not the internal framework:

In this setup, accepting a routine request is not the same as finishing a report, and finishing a report is not the same as delivering a Teams message. By keeping delivery receipts durable and code-owned, the workflow prevents a model from choosing recipients or blindly replaying messages when an acknowledgment drops.

Section Summary: Routines, Tool Search, and A2A are now GA, while Insights, Prompt-Agent Toolboxes, Voice Agents, and Long-Running Resilience enter public preview. Production workflows should pair scoped tool discovery with explicit protocol boundaries (Responses, Invocations, Activity) and code-enforced delivery guardrails.

Local AI and Windows Edge: Foundry Local 2.1.0 and Azure Local​

For Windows developers building on-device AI experiences or hybrid edge deployments, Foundry Local 2.1.0 shipped on September 30, 2026. Based on the official microsoft/foundry-local GitHub release notes and the Foundry roundup, version 2.1.0 removes one of the biggest friction points in local model testing: file duplication.

Bring Your Own Model (BYOM) Without Copying Files​

Previously, testing local weights often meant copying multi-gigabyte model folders into a managed cache directory. Foundry Local 2.1.0 lets developers register and unregister compatible ONNX Runtime GenAI models (upgraded to 0.17.1 alongside ONNX Runtime 1.30.0 in the early October CLI preview) directly in the local catalog without copying, downloading, or transferring ownership of the files on disk.

• Cross-Language Support: BYOM registration works across C++, C#, Python, JavaScript/TypeScript, and Rust using typed builders to supply model identity and task metadata.
• Unified Discovery & Tool Calling: Once registered, custom models work with the exact same discovery, loading, and session APIs as official catalog models. Version 2.1.0 also introduces text-based custom tools (including raw text payloads for code-patching and editing), stronger automatic Qwen tool calling for coding workflows, better recovery from malformed tool calls, and improved keyboard/screen-reader accessibility in the model browser.
• Important 2.x Migration Context: If your Windows or cross-platform application is upgrading from the 1.2.x branch through 2.0.1 (released September 1) to 2.1.0, note that in-process OpenAI-style SDK clients (ChatClient, EmbeddingClient, AudioClient) are deprecated and scheduled for removal at the end of 2026. Developers should migrate in-process code to the unified Session API (ChatSession, EmbeddingsSession, AudioSession). Furthermore, separate -winml add-on packages are no longer needed; the unified 2.x package automatically detects hardware and loads WinML, WebGPU, CPU, or CUDA execution providers. (Local HTTP REST endpoints remain OpenAI-compatible.)

Azure Local Extension 2609 (Preview)​

Separate from the desktop SDK, Foundry Local on Azure Local released extension 2609 (preview) in September. It enables Kubernetes service-account-token authentication by default for in-cluster inference, utilizing short-lived, audience-bound tokens and Kubernetes RBAC. Note that this only covers in-cluster inference—not control-plane requests—and administrators must verify their service account setup, including ensuring the extension managed identity holds the Reader role at the Arc-enabled cluster scope before upgrading.

Section Summary: Foundry Local 2.1.0 enables zero-copy registration of compatible ONNX Runtime GenAI models across five languages, while the 2.x branch consolidates WinML hardware acceleration into unified packages and deprecates in-process OpenAI-style clients in favor of typed Session APIs.

Speech and Content Understanding Updates​

Document extraction and voice pipelines received targeted developer tooling and service-side upgrades during September:

• Content Understanding Toolkit and CU CLI (Preview): Adds a preview command-line interface, toolkit, and analyzer schema references so developers can analyze local files and create, test, and manage custom analyzers locally rather than clicking through a web portal for every test. Remember that running file analysis and provisioning resources still incur cloud charges, and using the preview CLI does not alter the maturity level of the underlying API version you target.
• Speech LLM 2607 (Announced September 10): A service-side update that improves multilingual and mixed-language speech recognition and introduces a phraseList field to boost domain-specific terminology. Because it rolls out on the service side, customers do not need to swap out managed models, though Microsoft cautions that 2607 is a build identifier (not a launch date) and developers should test their specific fast-transcription or realtime paths rather than assuming identical behavior across both.
• Speech SDK 1.52: Adds preview commit support for PushAudioInputStream and allows developers to select a specific streaming transcription model. For JavaScript developers, 1.52 resolves bugs involving audio offsets on reconnect and missing channel metadata during conversation transcription.

Developer Tools, SDK Migrations, and CLI Breaking Changes​

If you live inside Visual Studio Code, Windows terminal environments, or CI/CD pipelines, September brought both quality-of-life installers and several breaking SDK/CLI changes that require attention before you run an upgrade script.

Foundry Dev Pack and Foundry Toolkit for VS Code​

Announced on September 15, the Foundry Dev Pack packages Foundry environment setup into a streamlined installer. Meanwhile, the Foundry Toolkit logged four rapid-fire releases throughout the month:

• 1.6.11 (September 2): Added hosted-agent validation support and Inspector event search/export.
• 1.6.12 (September 9): Added opt-in remote images, Windows profiling, and connection/replay diagnostics.
• 1.6.13 (September 16): Added private non-ACR container registry deployment UI, Fara AMD NPU support, and adapter/WinML tooling updates.
• 1.6.14 (September 23): Added solution templates, skill-only Toolboxes, steer/stop controls for active Responses work, and reconnection support for unfinished background responses.

Agent Framework and Language SDK Releases​

In the Microsoft Agent Framework, a September 4 update added a Python preview context provider for Azure Cosmos DB memory integration (supporting background extraction and cross-session recall; Microsoft warns developers to enforce authenticated user scoping rather than trusting a user ID passed inside a prompt). Python AG-UI support also reached stable status, whereas .NET support remains in preview. Core framework artifacts shipped across Python (1.17.0, 1.18.0, 1.19.0) and.NET (1.21.0, 1.22.0), adding vector-store integrations, checkpoint-restoration controls, and per-tool execution controls.

Across the Foundry Projects SDKs, keep an eye on these September package updates and breaking changes:

• Python (azure-ai-projects 2.6.0, 2.6.1, 2.7.0): Adds Insights and routine dispatch identity, polling fixes, preview Voice APIs, prompt harness/skills, and toolbox metadata. Check beta constructor and sample-limit changes. Additionally, in the Python evaluation patches (1.18.5 through 1.18.7), upgrading to 1.18.7 removes the aoai_output_items_page_size option—remove that parameter from your code before updating.
• JavaScript/TypeScript (@azure/ai-projects 2.6.0, 2.7.0, 2.7.1): Adds Insights and preview Voice APIs, relocates beta.voiceAgents operations, and fixes browser bundling in 2.7.1.
• .NET (Azure.AI.Projects 3.0.0-beta.2 and beta.3): Remains on a beta track, adding Insights, realtime-session startup, and preview telephony/conversation clients. For retrieval code, Azure.Search.Documents 12.1.0-beta.3 (September 14) changes the streamed-retrieve protocol overload to AsyncStreamingResult<SseItem<BinaryData>> and restores indexer create/update parameters.
• Java (azure-ai-projects 2.5.0, 2.6.0, 2.6.1; azure-ai-agents 2.6.0): Adds Insights, routine authorization, tracing/output controls, preview Voice clients, and serialization fixes. Verify required authorization arguments.
• Hosted-Agent Runtime Packages: Review AgentServer Core 2.2.0, Responses 2.2.0 (and 2.3.0 beta), and Invocations 1.2.0. Crucially, AgentServer Core defaults to 100% Azure Monitor trace sampling—enterprise admins should review trace volume and telemetry egress costs before rolling it into high-traffic production.

Azure Developer CLI (azd) Breaking Recipe Changes​

Alongside core azd releases 1.33.0 through 1.34.2, the Foundry CLI extensions introduced two breaking changes for teams using summer deployment scripts:

1. azure.ai.agents beta.18 (September 30): Requires azd 1.34.2 or later. Runtime definitions must now reside directly inside azure.yaml or via an explicit root $ref; legacy AgentManifest/config discovery is rejected. Long-running invocations also move to the --long-running flag and the invocations command group.
2. Connections and Toolboxes Extensions: Removed standalone deploy commands in favor of azd deploy <service>.

Pricing Policy Shifts and Model Retirements​

Finally, FinOps teams and cloud architects need to account for deployment pricing changes that took effect on September 1, 2026 (originally announced in July):

• Standard Deployment Premiums: Apply only to models launched on or after September 1, 2026; continuing to run an existing Standard model does not trigger the new premium.
• EU Data Zone and Non-US Regional PTUs: Follow a stricter rule—pricing increases apply to existing customers as well. Global deployment pricing remains unchanged by this policy, and the APAC Data Zone is now generally available under the updated structure.

Also check your deployment inventories against two September lifecycle retirement dates:

• September 15, 2026 (MAI-Transcribe-1, version 2026-01-23): Scheduled retirement deadline; inventory existing deployments and evaluate MAI-Transcribe-1.5.
• September 24, 2026 (gpt-chat-latest, snapshot 2026-06-24): Scheduled deadline for this specific pinned snapshot (this does not retire later snapshots or the rolling alias itself).

Practical Next Steps for Engineering Teams​

1. Audit azd and Python Evaluation Scripts First: Before pulling September dependencies into CI/CD, remove aoai_output_items_page_size if upgrading to Python evaluation 1.18.7, and update azure.yaml files if adopting azure.ai.agents beta.18.
2. Test Routines in Manual "No-Send" Mode: When adopting GA Routines for automated reporting, validate the collector agent and trace output manually before enabling calendar schedules or Teams Activity delivery.
3. Check AgentServer Core Trace Sampling: If deploying Hosted Agents with AgentServer Core 2.2.0, adjust the default 100% Azure Monitor trace sampling rate to match your observability budget.