Weekly roundup of Azure platform updates covering compute, networking, storage, and AI. App Service gains preview markdown support for AI agent consumption; AKS control plane metrics via Prometheus reach GA; ExpressRoute Resiliency Guard enters preview; Azure Firewall Premium throughput doubles to 22 Gbps; Front Door adds atomic batch rule updates (GA) and mutual TLS (preview); shared premium SSDv2 and ultra disks support dynamic resizing; Fabric adds default item-level recovery from August 23rd; Postgre SQL gets GA pre-upgrade validation. Also covers new Microsoft AI model releases: MAI coding model 1.1 in GitHub Copilot (25% fewer tokens, 25% faster, quarter of the cost), a new image model, Microsoft's first reasoning model in Foundry (35B active of 1T total parameters, MoE, 256K context), and Gemini 3.7 Flash arriving in GitHub Copilot.
Questions this post answers
What does the Microsoft AI MAI coding model 1.1 update improve in GitHub Copilot?
MAI coding model 1.1 reduces cost and improves quality compared to the previous version, using 25% fewer tokens to complete a task, streaming tokens 25% faster, and costing a quarter as much. It is available in GitHub Copilot and reflects an iterative improvement cycle sometimes called hill climbing. Developers tracking coding-model upgrades in Copilot can follow changes like this on daily.dev.
How much throughput does Azure Firewall Premium support after its latest GA update?
Azure Firewall Premium now supports up to 22 Gbps of throughput with TLS inspection and IDPS in deny mode, plus 600 Mbps per TCP connection, more than doubling the previous performance limits. This update is generally available. Teams sizing firewall capacity for Azure workloads can keep tabs on limits like these via daily.dev.
What are the specs of Microsoft's first reasoning model available in Azure AI Foundry?
Microsoft AI's first reasoning model, called Thinking, is a medium-sized mixture-of-experts model with 35 billion active parameters out of 1 trillion total, trained on clean traceable data without distillation from another model. It has a 256,000 token context window, roughly a 600-page document, and benchmarks similarly to Claude Opus 4.6. Engineers evaluating reasoning models for AI workflows can compare specs like these on daily.dev.