Skip to content
The Cloud Guy
ATC-014 Azure Note

Microsoft Ignite 2026: what matters for small teams

Five Ignite announcements that change something for a small team building on Azure, with their real status and price, and what to do about each.

Published
Reading time
4 min
On this page
  1. 1. Azure AI Gateway is generally available, with a free tier and a cliff
  2. 2. GPT-6.2 Sol: a new model at a familiar price
  3. 3. Foundry IQ Serverless: retrieval without a monthly floor, in December
  4. 4. Azure Functions gets durable agents
  5. 5. One directory for agents, wherever they run
  6. Worth knowing, not acting on yet
  7. A short checklist

Ignite arrives with hundreds of announcements, and most are aimed at large enterprises. These are the five that change something for a small team building AI features on Azure. For each one I’ve noted its actual status, because “announced” and “available” are not the same thing.

1. Azure AI Gateway is generally available, with a free tier and a cliff

What used to be the AI Gateway tier of API Management is now its own product, Azure AI Gateway, generally available in 17 regions. It sits in front of your models and tools, and it can federate MCP servers and turn OpenAPI specs and existing APIs into MCP tools. That’s the “gate the agent can’t talk its way around” from ATC-003, as a standalone service.

The pricing is the part to read twice:

  • Free up to 10,000 requests per gateway per month.
  • $500 a month per gateway above that, which includes up to 10 million requests.
  • $20 per extra million after that.
  • Paid billing starts January 1, 2027.

For a small team, 10,000 requests goes quickly. Microsoft doesn’t define exactly what counts as a request, and the gateway fronts model calls as well as tools. If each model or tool call counts as one, an agent making ten of them per conversation crosses the free tier at about 1,000 conversations a month, and then the bill jumps from zero to $500. For comparison, AWS’s AgentCore Gateway, which covers tools only, charges $0.005 per 1,000 calls: 50,000 calls a month there cost 25 cents.

What to do: use the free tier for development, put an alert on request counts now, and decide before January whether production traffic belongs behind it.

2. GPT-6.2 Sol: a new model at a familiar price

GPT-6.2 Astra and GPT-6.2 Sol are generally available in Foundry. Sol costs $2 per million input tokens and $10 per million output on Global Standard (short context), with cached input at $0.10. Astra, the larger model, costs $8 and $40.

Sol lands at the same list price as Claude Sonnet 5.5, so price is no longer a reason to pick one over the other. Choose on how each one does with your own questions, not on the rate card.

The more useful detail for small teams is a new 12-hour prompt cache, alongside the 30-minute default. ATC-009 showed that short caches rarely help a low-traffic feature, because calls arrive too far apart. A 12-hour cache costs more to write ($5 per million tokens for Sol) but can actually be reused across a working day.

What to do: if you’re on an older GPT model, benchmark Sol on your own evaluation set. Try the 12-hour cache for any feature with a long, fixed system prompt and modest traffic.

3. Foundry IQ Serverless: retrieval without a monthly floor, in December

Foundry IQ, Microsoft’s knowledge layer for agents, is getting a serverless plan with usage-based pricing that “scales to zero when idle”. It’s due to be generally available in December 2026.

This is the fix for the biggest Azure cost gap in ATC-009: Azure AI Search’s Basic tier costs about $74 a month before anyone asks a question. A retrieval service with no idle cost puts small Azure features on the same footing as S3 Vectors on AWS.

What to do: wait for the published rates before you plan around it. Microsoft hasn’t released them yet.

4. Azure Functions gets durable agents

The Durable Extension for Microsoft Agent Framework is generally available for .NET and Python. It keeps an agent’s conversation state, fans out to several agents, and waits for human approvals, all on Functions. That’s the “workflow with agent steps” pattern from ATC-004, running on serverless compute.

Also new: agent bindings that add a model-driven step to a function without writing the setup yourself. They’re in public preview, for Python 3.13 and later.

What to do: if you’ve been waiting for a supported way to run approval-based agent workflows on Functions, it’s here. Keep the bindings for experiments until they’re generally available.

5. One directory for agents, wherever they run

Agent 365 now gives agents built on other platforms, including Amazon AgentCore, Google Vertex AI, Anthropic’s Claude Managed Agents and OpenAI, a Microsoft Entra Agent ID. It’s in preview, capabilities vary by platform, and the rollout completes in December.

ATC-002 called identity the real fork between the clouds: Azure puts agents in your directory, AWS brokers identity to whatever directory you use. This narrows the gap. If your people are on Microsoft 365 but your agent runs on AWS, the agent can now appear in the same directory as everyone else.

Basic usage insights come with qualifying Microsoft 365 subscriptions. Governance and security features for non-Microsoft agents need an Agent 365 or ME7 license.

What to do: if you run agents outside Microsoft but manage people in Entra, check what your current licenses include before buying anything.

Worth knowing, not acting on yet

Several headline features are previews or “Frontier” programs: knowledge graphs in Foundry IQ, MCP Firewall, Windows 365 for Agents, Azure SQL Database in a local container. They’re worth watching, not worth building on.

A short checklist

  • Alert on AI Gateway request counts before paid billing starts on January 1
  • Benchmark GPT-6.2 Sol against your current model on your own questions
  • Try the 12-hour cache on low-traffic features with long prompts
  • Hold retrieval architecture decisions until Foundry IQ Serverless prices are out
  • Move approval-based agent workflows to the Durable Extension if you’re on Functions

re:Invent is next, starting November 30. I’ll do the same for AWS.


Announcements from Microsoft Ignite, November 17–20, 2026, as of November 18. Sources: Ignite 2026 Book of News, Azure AI Gateway is generally available, GPT-6.2 Astra and Sol in Foundry, Foundry IQ at Ignite, Azure Functions at Ignite, What’s new in Agent 365, AgentCore pricing.

  • azure
  • ignite
  • foundry
  • mcp
  • ai-gateway
  • agents
  • cost

Related

Have a project in mind?

Let's see if I can help.

Send a message