Agent 365 Is $15/User. The 4 Models Powering It Aren't All Microsoft's.

13 min read By AI Skills Directory

Agent 365 ships Friday at $15/user. After the OpenAI restructure, here are the 4 models — GPT-5.5, MAI, Claude, open-weights — and what each costs you.

Table of Contents

Microsoft Agent 365 goes generally available on Friday, May 1. The pricing is settled: $15 per user per month standalone, or bundled into the new $99-per-user “Frontier Worker Suite” (E7) alongside Microsoft 365 E5, Copilot, and Entra Suite. What’s not settled — and what every IT buyer in an EA renewal cycle is asking — is which model actually powers Agent 365 on Day 1, and what shifts under the hood by Q3.

That question got dramatically more interesting two days ago. On April 27, Microsoft and OpenAI restructured their partnership, ending Microsoft’s exclusivity on OpenAI’s IP. Microsoft keeps a license through 2032, but it’s no longer exclusive — and OpenAI is now free to sell its models on AWS and Google Cloud. Microsoft, in the same breath, is openly pushing a multi-model story: Claude in Copilot via Foundry, in-house MAI models in preview, and “without locking customers in” as the official talking point.

So the procurement question isn’t “is GPT-5.5 inside Agent 365?” It’s: which of the four model paths your $15/user license is buying access to over the next 18 months — and what each one means for governance, data residency, and total cost.

Here’s the four-scenario read for IT buyers reviewing Agent 365 ahead of Friday’s GA.

What Agent 365 Actually Is (in 90 Seconds)

Agent 365 is Microsoft’s control plane for AI agents — the centralized layer your IT, security, and business teams use to observe, govern, and secure AI agents across the enterprise. Picture how Microsoft Defender governs identities and devices today; Agent 365 is the same idea applied to AI agents (Copilot agents, custom Copilot Studio agents, agents built in Microsoft Agent Framework, third-party agents wired in via standards).

It does four things in plain English:

  1. Inventory — every agent in your tenant, who built it, who can use it
  2. Govern — policies for what agents can access (data, apps, identities)
  3. Observe — logs of what agents did, when, and on whose behalf
  4. Secure — alerts when an agent acts outside its policy or gets compromised

It does not ship as the agent itself. Agent 365 governs the agents your tenant runs. Those agents — Copilot, Copilot Studio agents, partner agents — are what actually run the model. Which is exactly why the model question matters.

The Pricing Math IT Buyers Are Running Friday

Per the SAMexpert E7 breakdown and Microsoft’s own March 9 announcement:

SKUPrice (annual commit)What’s in it
Microsoft 365 E5$60/user/mo (rising July 2026)Existing E5 — productivity + security
Microsoft 365 Copilot$30/user/moThe Copilot you already buy
Entra Suite$12/user/moIdentity governance + verified ID
Agent 365 (standalone)$15/user/moAgent control plane — new
E7 “Frontier Worker Suite” (bundle)$99/user/moAll four above
Same components unbundled$117/user/mo—

The bundle saves 15% versus buying à la carte. Microsoft is also offering promotional pricing of 10% off for 10+ seats and 15% off for 100+ seats during the GA window. For a 1,000-seat enterprise, that’s a $1.18M annual difference ($1.41M unbundled vs $1.18M E7 with promo) — enough to make the procurement-team math itself worth a Friday morning.

But the math is only one input. The model question is the other.

What Changed on April 27 (and Why It Reshapes the Agent 365 Question)

Per Microsoft’s own blog and corroborating Bloomberg coverage, the restructured deal does five things:

  1. Microsoft’s exclusivity on OpenAI’s IP ends. OpenAI is free to sell on AWS and Google Cloud (and reportedly is talking to both).
  2. Microsoft keeps a license through 2032. Non-exclusive, but durable.
  3. OpenAI products ship first on Azure “unless Microsoft cannot or chooses not to support the necessary capabilities” — a clause that quietly lets Microsoft opt out of bringing certain OpenAI products to Azure (e.g., a hypothetical OpenAI hardware or specialty model Microsoft doesn’t want on its cloud).
  4. Revenue-share is capped and runs through 2030.
  5. The “AGI determination” clause is gone. Microsoft no longer needs to react if OpenAI declares it has reached AGI.

For Agent 365 specifically, the third change is the load-bearing one. Microsoft has, for the past 12 months, publicly broadened its model story — adding Anthropic to Copilot Studio in September 2025, putting Claude Opus 4.6 on Microsoft Foundry, shipping in-house MAI-1 in preview. The April 27 restructure is the public-facing unlock. From May 1 forward, Microsoft can credibly tell enterprise IT: “You’re not locked to OpenAI. You’re locked to us, and we’ll route you to the best model for your job.”

Here’s what that means in practice. The four model paths an Agent 365-managed agent will likely run on between May 1 and Q3.

Scenario A: GPT-5.5 / OpenAI (the Day-1 default)

Most likely on May 1. Copilot, the agent most enterprises will have Agent 365 governing first, has run on OpenAI models since 2023. Wave 3 of Microsoft 365 Copilot ships the same day with “expanded model diversity from both OpenAI and Anthropic” — but the default for net-new agents and the existing Copilot install base remains OpenAI’s frontier model.

What this means for procurement:

  • Existing Copilot tenants don’t see a model change Friday morning
  • OpenAI’s 2032 non-exclusive license is enough runway for normal renewal cycles
  • Data residency: stays in Azure tenants where Copilot was already running — unchanged
  • Cost trajectory: OpenAI’s per-token pricing drops on a quarterly cadence; bundled into Copilot, no separate line item
  • Risk: OpenAI launching a parity model on AWS in Q3 means your competitors could run the same model in a different cloud while you stay in Azure. Not a 2026 problem; possibly a 2027 negotiation lever.

Scenario B: Microsoft MAI (the in-house hedge)

Microsoft is testing MAI-1-preview, its in-house frontier model, public reporting tied this directly to the April 27 restructure as the strategic context. Microsoft hasn’t published an Agent 365 launch date for MAI as a default model — but the talking points around April 27 (“operating openly across clouds and data services without locking customers in”) are pointing toward MAI showing up as an option in Copilot Studio first, possibly Q3 2026.

What this means for procurement:

  • A Microsoft-stack-native model with no third-party licensing fee passed to you
  • Data residency: cleanest of the four scenarios — MAI runs in Azure, trained by Microsoft, governed by Microsoft contracts you already have
  • Cost trajectory: this is the model Microsoft will eventually use to push back on OpenAI’s pricing. Expect it to be the cheap option
  • Risk: capability gap vs OpenAI/Anthropic frontier models. MAI-1-preview hasn’t been benchmarked publicly. Don’t bet your most demanding agents on it for at least two release cycles.
  • Tells: watch for Microsoft to start using MAI as the default for narrow, cheap agents (translation, summarization, formatting) before any high-stakes reasoning agent

Scenario C: Anthropic Claude via Microsoft Foundry

Already shipping. Claude is available in mainline Microsoft 365 Copilot chat via the Frontier program, and Claude Opus 4.6 is on Microsoft Foundry on Azure as of late 2025. Microsoft Foundry is currently the only cloud where you can access both Claude and GPT frontier models on the same platform — a procurement reality that genuinely matters for enterprises that don’t want a multi-cloud AI strategy just to run two models.

What this means for procurement:

  • Your Copilot Studio team can already pick Claude Sonnet 4 or Opus 4.1 for new agents per Microsoft’s Copilot Studio multi-model docs
  • Data residency: Claude-via-Foundry runs in Azure regions, governed by Microsoft contracts (not Anthropic’s direct contracts) — important if your legal team has spent six months negotiating Microsoft’s terms and doesn’t want to layer a second vendor’s
  • Cost trajectory: Microsoft is reportedly subsidizing Claude usage in Copilot to compete with native Claude.ai for enterprise deals — pricing may shift, but the direction is “competitive vs OpenAI”
  • Use case: the Researcher agent in Copilot 365 already runs on Claude. Long-context, complex-reasoning, code-generation tasks are the natural Claude lane.
  • Tells: watch for which Copilot agents Microsoft markets as “powered by Anthropic” vs “powered by OpenAI” — that’s Microsoft’s signal of which model fits which job

Scenario D: Open-Weights (DeepSeek, Qwen, Llama)

Theoretical today, increasingly real. Microsoft hasn’t announced open-weights model support for Agent 365-governed agents. But: Microsoft Foundry already hosts open-weights models for Azure customers; Copilot Studio’s multi-model framework is provider-agnostic; and the pressure on Microsoft to not be locked to two American vendors (OpenAI + Anthropic) is real, especially for European public-sector buyers. Adding DeepSeek V4 or Qwen 3.6 to the Copilot Studio model picker in 2026 would be a small engineering step and a large procurement signal.

What this means for procurement:

  • Cheapest of the four paths — open-weights inference on Azure runs at infrastructure cost only
  • Data residency: highest control — you can pin the model to a specific Azure region, no model-vendor contract
  • Risk: capability + support. Open-weights models lag frontier on hard reasoning, and Microsoft Support’s response pattern for open-weights agent failures is undefined.
  • Sovereignty angle: this is the path that German Mittelstand buyers and EU AI Act-regulated industries will quietly push for over the next 12 months. Watch for a Microsoft announcement in Q3 that adds at least one open-weights model to the Copilot Studio picker.

The 4-Question Decision Framework for Friday

If you’re an IT decision-maker reviewing Agent 365 in your EA renewal package this week, the four questions to bring to your Microsoft account team:

1. “Which model is the default for new agents created in Copilot Studio after May 1?” The answer should be specific (e.g., “GPT-5.5 for new Copilot Studio agents; you select Claude or other models per agent”). If the answer is vague, push for the agent-creation flow walkthrough on a sandbox tenant.

2. “What’s the licensing path if we want to standardize on Claude via Foundry across our Copilot Studio agents in Q3 2026?” This tests whether Microsoft will charge incrementally for non-default model use, or whether the $15/user Agent 365 license already covers cross-model agents. Get it in writing.

3. “What’s the data-residency commitment for Claude-via-Foundry vs OpenAI vs MAI?” All three should run in Azure tenant regions, but the contracts behind each are different (Microsoft’s standard Azure contract for OpenAI/MAI, a Microsoft-Anthropic-flow-through arrangement for Claude). Your legal team will want to read the actual data-processing agreement language for each.

4. “What’s Microsoft’s commitment on adding at least one open-weights model to the Copilot Studio picker by end of 2026?” This is the diversification test. Microsoft’s answer signals how seriously they take the European-sovereignty buyer cohort and how much pricing pressure they expect from open-weights in 2027.

If your account team has clean answers to all four by Friday, Agent 365 is a normal SKU upgrade. If three are vague, Agent 365 is a multi-quarter procurement project that should defer to Q3 review.

Who Should Buy Agent 365 on Day 1

If you already run Microsoft 365 E5 + Copilot: the E7 bundle is a defensible upgrade for the 15% discount alone, even before counting Agent 365’s governance value. Lock in promotional pricing during the GA window.

If you run E3/E5 without Copilot: Agent 365 standalone at $15/user is rational only if you have actual agents to govern. If your Copilot adoption is below 30% of seats, the agent-governance load isn’t there yet. Defer.

If you’re evaluating Agent 365 vs Gemini Enterprise vs Claude Cowork on Mobile: different product categories. Agent 365 governs agents you’ve already built or bought. Gemini Enterprise + Cowork are agent-creation surfaces. Most large enterprises will end up with Agent 365 plus one or two agent platforms, not Agent 365 vs one of them.

If you’re a Microsoft-shop CISO worried about agent governance: Agent 365 is the only product on the May 1 enterprise-AI release calendar that addresses your specific role. The competition is third-party (Sentra, Cyera, AI agent governance startups) and not yet at Microsoft-tenant integration depth.

What This Means for You

If you’re an IT decision-maker in an EA renewal cycle: the Agent 365 / E7 procurement decision is one of the larger 2026 line items. The April 27 OpenAI restructure means the model story will keep moving — your contract should be flexible enough to accommodate the Q3 multi-model roll-out.

If you’re a Microsoft enterprise architect: the four-scenario framework above is the conversation to have with your stakeholders this week. Default to GPT-5.5, but design new Copilot Studio agents with the model-picker abstraction so a Q3 swap is a config change, not a rebuild.

If you’re a security / compliance lead: the Agent 365 control plane is genuinely the missing piece in your AI agent governance stack. The data-residency questions on Claude-via-Foundry vs OpenAI are the load-bearing concern; everything else is configuration.

If you’re a CFO reviewing the $99 E7 bundle: the math is real. 15% bundle discount + 10-15% volume promo during the GA window is the cheapest moment to lock in. Renewal pricing in 2027 will not be friendlier.

The bottom line: Agent 365 ships as a control plane, not as an agent. The model question — which of GPT-5.5, MAI, Claude, or open-weights powers the agents you run under it — is the actual procurement decision for the next 18 months. Friday’s GA settles the SKU. The model strategy stays open through at least Q3 2026.

Sources:

Build Real AI Skills

Step-by-step courses with quizzes and certificates for your resume