Microsoft’s September 2026 Foundry updates include several generally available capabilities and a set of public previews. Routines, Tool Search and Agent-to-Agent are GA, while Insights, voice agents, long-running hosted-agent resilience and additional Model Router controls are in preview. The company also expanded the GPT-6 family and updated local, speech, content-understanding and developer tooling.

GPT-6 Astra, GPT-6 Luna and GPT-6.1 Sol are generally available through supported serving options. The models have 1.05 million-token context windows and support up to 128,000 output tokens within a shared budget, along with Responses, Chat Completions, streaming, function calling, structured outputs and image input with text output. Astra and GPT-6.1 Sol have an April 2026 knowledge cutoff, while Luna’s is May 2026; GPT-6.1 Sol tool calls require Responses.

Model Router adds preview session affinity and request-level routing metadata, and expands to 32 regions. Session affinity prefers the same eligible model across related Chat Completions requests but does not guarantee model reuse or a cache hit, because safety, capability, quota, availability and fallback requirements still apply. The returned metadata can show routing and fallback behavior.

Routines are now GA and can start one agent from a timer, recurring schedule, GitHub issue event or supported Teams channel event, with configuration and run history retained in a Foundry project. They do not replace multi-agent orchestration or approval workflows, and the agent-scheduled reminder tool remains in preview. Tool Search is also GA, allowing agents to discover and invoke tools on demand through tool_search and call_tool instead of placing every definition in each request.

Insights enters public preview as a trace-based review feature that groups recurring behavior into findings with supporting evidence, likely causes and proposed next steps. The article advises inspecting linked traces and testing changes against the same input before applying them. Voice agents and long-running hosted-agent resilience are also previews; resilience still requires documented server configuration, checkpointing and idempotency because reconnection does not make repeated external actions safe.

Speech and content-understanding updates include a preview Content Understanding Toolkit and CU CLI for local-file analysis and analyzer management. Speech LLM 2607 improves multilingual and mixed-language recognition and adds a phraseList field, while Speech SDK 1.52 adds preview PushAudioInputStream commit support, streaming-model selection and JavaScript reconnect fixes. The article notes that rollout timing and capability coverage can differ across transcription modes.

Foundry Local 2.1 supports registering compatible ONNX Runtime GenAI models already on disk without copying or transferring their files, with support across C++, C#, Python, JavaScript/TypeScript and Rust. It also adds custom text tools and improvements to cancellation, token accounting and model-browser accessibility. The Azure Local extension 2609 separately adds default Kubernetes service-account-token authentication for in-cluster inference, but not for control-plane requests.

Developer tooling updates include the Dev Pack installer, Foundry Toolkit templates and controls, Agent Framework memory and interactive-experience changes, SDK releases, Azure Developer CLI updates and beta Foundry extensions. The article also highlights September pricing-policy and model-retirement notices, emphasizing that pricing depends on model, geography and reservation terms and that listed retirement dates are schedules rather than observed shutdowns. It directs users to current documentation for migration, availability and preview boundaries.