The AI Wire

claude.com ▸ tooling· filed 28 Jul 2026

MCP Version 5 Removes Stateful Sessions, Adds OAuth and Extensions Framework

Anthropic released version 5 of the Model Context Protocol, replacing its stateful session model with a stateless request/response architecture. The spec adds OAuth 2.0 authorization and a versioned extensions framework for features like interactive UIs and long-running tasks. Adoption data shows over 400 million monthly SDK downloads and 950 MCP servers listed in Claude’s directory.

azure ▸ models· filed 28 Jul 2026

Kimi K3: 2.8 Trillion-Parameter Model with Million-Token Window Now Available on Microsoft Foundry

Moonshot AI's Kimi K3, a 2.8 trillion-parameter open-weight model with a 1 million-token context window, is available for deployment on Microsoft Foundry through Fireworks AI. The model's large context window eliminates the need for chunking when handling long documents or codebases, though attention reliability across the full window remains an active area of research. Pricing is $3.30 per million input tokens and $16.50 per million output tokens, reflecting the model's size and inference cost.

azure ▸ tooling· filed 28 Jul 2026

Two Layers for GitHub Copilot Cost Governance as the Promotional Cushion Nears Its End

GitHub Copilot's promotional credit uplift ends September 1, 2026, shifting focus to spend governance. Native budget controls provide financial caps across all surfaces, while a second layer using Azure API Management enables real-time token quotas for IDE and CLI usage. Organizations should enable native budgets now and consider piloting the more granular BYOK approach if needed.

azure ▸ product· filed 27 Jul 2026

Azure Databricks Adds Claude Opus 5 to Model Serving

Azure Databricks has made Claude Opus 5 generally available via its AI Model Serving platform, allowing organizations to use the latest Anthropic model for inference within the Databricks environment. The model targets advanced reasoning and agentic workflows, but details on performance, pricing, and differentiation from its predecessor remain scarce, leaving developers to conduct their own testing.

azure ▸ azure· filed 27 Jul 2026

Microsoft Releases AI Gateway Tier for Azure API Management

Microsoft has released a public preview of the AI Gateway tier for Azure API Management, offering a dedicated pricing and feature tier for routing and securing AI API calls. The announcement is short on specifics about features, performance, or region availability. Engineers can test the tier in staging, but its value depends on whether it provides distinct capabilities beyond existing API Management tiers.

devblogs.microsoft.com ▸ tooling· filed 26 Jul 2026

Microsoft Releases Stable Agent Harness for Python and .NET

Microsoft has shipped a stable agent harness for Python and .NET as part of its Agent Framework, providing runtime scaffolding for turning language models into agents with features like function invocation, memory, and approval workflows. The harness is customizable and built for long-running tasks, though some opt-in features are withheld pending customer feedback.

azure ▸ product· filed 24 Jul 2026

Microsoft Drops One-Line Memory for Agent Framework — No Orchestration Code Required

Microsoft's Agent Framework now includes a CosmosMemoryContextProvider that provides one-line memory integration via Azure Cosmos DB. It handles extraction, storage, and retrieval of memories without requiring custom orchestration code. The preview package is Python-only and uses background processing for non-blocking memory updates.