GitHub Copilot doubled the multiplier on Claude Opus 4.7 last Friday. From 7.5x to 15x. Same multiplier line, twice the cost. Pro plan customers had already lost Opus access entirely on April 20. Same Friday, Microsoft Agent 365 went GA at $15 per user per month. Same day. Victor Waenerlund’s three-sentence read of those events on LinkedIn was the only post in my feed that named the play: “Microsoft has stopped trying to own the IDE or the model. They want a tax on the model you choose, and an audit log on the agents you run.” Three sentences. Everything that follows in this piece is detail.

What changed in 24 hours

Two events, same day, ordered:

• The Opus 4.7 multiplier on GitHub Copilot moved from 7.5x to 15x for Pro+, Business, and Enterprise. The change is buried in the editor’s note on the April 16 Copilot changelog, dated May 1. The Pro plan had lost Opus completely on April 20.
• Microsoft Agent 365 entered general availability at $15 per user per month, or bundled in Microsoft 365 E7. Microsoft’s own security blog post describes it without ceremony: a control plane that “observes, governs, and secures agents” no matter where they run. Microsoft Foundry, AWS Bedrock, Google Gemini Enterprise Agent Platform, or local agents like OpenClaw, Claude Code, and GitHub Copilot CLI.

Microsoft is not blocking you from using Anthropic. They are charging you twice for using it. Once in the multiplier on the IDE that calls the model. Once again in the control plane that audits the agents that IDE writes.

The license clause that nobody quoted

The detail no thread quoted on May 1 lives inside the same Microsoft blog post, near the bottom, where it explains the license model. The exact wording: “Each Agent 365 license covers an individual who manages or sponsors agents, or uses agents to do work on their behalf.” That last clause is the one that matters.

The price does not scale by agent. It scales by agent shadow. If your organization has 800 people and each of them has an agent doing something on their behalf, that is $12,000 per month. $144,000 per year. Without buying a single token from Anthropic, OpenAI, or Google. That is just the toll for showing the audit log of what those agents did.

The list of launch partners is harder to ignore. Accenture, KPMG, Deloitte, EY, PwC, Capgemini, Cognizant, Avanade, TCS, Slalom, Bechtle, Protiviti. Every one of them is a 20-year compliance shop. Microsoft just gave them a new SKU.

What the control plane actually does

The capabilities announced on May 1 are not theoretical. The list:

• Cross-cloud discovery. Registry sync with AWS Bedrock and Google Cloud (public preview). Microsoft inventories your agents on other vendors’ clouds.
• Local discovery. Defender and Intune detect agents running on managed endpoints. The detection list names OpenClaw, Claude Code, and GitHub Copilot CLI explicitly. If your engineers run Claude Code on a corporate laptop, Defender already has policy to block it.
• Governed runtime. Windows 365 for Agents, Cloud PCs purpose-built for agentic workloads, with Intune policies and runtime blocking.
• Network controls. Microsoft Entra extended network policies to Copilot Studio agents and to local agents like OpenClaw, filtering unsanctioned web destinations and blocking prompt-injection attempts.

The structure is elegant. Microsoft does not need to win the model fight (they are losing it to Anthropic on code, drawing with OpenAI on chat). They do not need to win the IDE fight (Cursor and Antigravity own that). What they need is for each of those vendors to keep selling inside enterprises that pay $15 per user per month to audit what those vendors did.

Vas at Varick Agents wrote the long version of why this matters. His April 29 essay (461K views) named the operator gap: 78% adoption, 80% no impact. Process is the bottleneck. Microsoft just shipped the audit-of-the-bottleneck as a $15-per-user-per-month SKU before most enterprises knew it was a SKU. The chorus around the essay had been pointing at the gap for two months. Microsoft was already filling it on a release schedule. The point of friction Vas described as “the operating layer underneath the model” turned out to be the SKU Microsoft was building behind closed doors.

Five times Microsoft has moved up a floor

I have been in computing for 36 years. I started in 1990, at 15, on a Commodore 64 and a Texas Instruments. Five times I have watched Microsoft do exactly this play. Each one looks the same: the layer below becomes a commodity, Microsoft does not fight to keep winning there, they move up one floor.

Cycle one: mid-1980s. When the BIOS and IBM-compatible hardware turned into commodities, Microsoft refused to compete on hardware. They went one floor up and held the OS instead. DOS, then Windows. The cheap layer was somebody else’s revenue. The layer Microsoft kept compounded for thirty years.

Cycle two: mid-1990s. WordPerfect, Lotus 1-2-3, and Harvard Graphics each owned a category outright. Microsoft never tried to win each fight in isolation. They bundled the three into Office and sold the suite. WordPerfect and Lotus died without ever losing a single feature war against the Microsoft equivalent. They lost the bundle war.

Cycle three: 1998 to 2003. Netscape, and then Mozilla, dominated browsers. Microsoft shipped Internet Explorer 6, which was inferior, and everyone knew it. The browser fight was lost on the merits. Microsoft won it anyway by bundling IE inside Windows. The win came at the OS layer, not at the browser layer.

Cycle four: 2017 to 2022. Slack defined enterprise chat. Teams was uglier (still is, in 2026), but Microsoft bundled it inside Microsoft 365 and let Slack be technically superior while taking 80% of the enterprise share. Slack ended up at Salesforce not because they lost the product fight, but because they lost the bundle fight.

Cycle five: 2026. Anthropic, OpenAI, Cursor, and Antigravity own the model and the IDE. Microsoft is not going to ship a better model. Microsoft is not going to ship a better IDE. They moved up a floor and shipped Agent 365. The model fight gets won at the audit log.

Five times. The compression is the new thing. The first four cycles took three to five years to consolidate. This one is 24 hours from GA and already has the partner catalog listed.

Five questions before the first Agent 365 invoice arrives

If your next executive committee meeting still has “AI strategy” framed around which model to choose, there are five places to push before the next budget decision.

Inventory comes first, and it is the question that produces the longest silence. How many agents run in your organization right now, on which cloud, against which data, with whose credentials? When nobody on the committee can answer in 15 minutes, somebody outside is going to sell you the answer in 15 days.

After that, exposure to the multiplier. The Opus multiplier on GitHub Copilot just went from 7.5x to 15x with no review meeting attached. The honest question is whether you have any alternative at all, or whether the price your engineers pay is whatever Microsoft sets next Friday.

The third place to push is abstraction, and it is the cheapest one to fix this quarter. If your access to Anthropic, OpenAI, and Google sits behind a layer your team controls, switching providers is an environment variable. If each team is going direct to the provider with its own wrapper, switching is a six-month project the next time a price doubles.

Position is the question most committees skip because it sounds abstract until it is not. When Microsoft, Google, or AWS launches the control plane that audits your agents on a competitor’s cloud, the answer to “are you the auditor or the audited” decides whether your governance is a product line or a line item.

Ownership closes the list. Whoever on the executive committee is named owner of the “agent control plane” problem should be able to answer the four above in 15 minutes. If the answer is “we need to come back to it,” the problem already has an owner, and it is not you.

If your team cannot answer all five with clarity, the next useful conversation runs about two hours. We map which agents you run, where the single-vendor dependencies are, and which parts of your model access remain configurable without a rewrite. No quote attached. The address is the usual one: info@iqsource.ai.

What we do at IQ Source about this

AI Maestro is the audit Microsoft is selling, run from the other side. The difference is position. Microsoft audits you to sell you the log SKU. AI Maestro audits you so that when the SKU arrives, you already have the map, you already know which agents run against which data, and you can decide whether the toll is worth paying. The audit you commission yourself is what gives you the ground to negotiate the audit Microsoft is going to run.

Tech Partner applies when the company looking at the control plane is not just a customer of the control plane but also a product. Software companies whose product lives in the critical zone from day one now have two distinct questions to answer. One, how to manage the dependency on a single model provider. Two, how not to end up being the supplier for the next version of Microsoft Agent 365. Both share tooling. Both require engineering that lives in the codebase, not in committee slides.

On April 23 I wrote about the other half of the playbook: OpenAI doubling the price of the model. That was the classic Microsoft 1995-to-2010 move, extraction inside a layer. What Microsoft made public on May 1 is the Microsoft 1985 move: moving up a floor. Both run at the same time. Both arrive on the same invoice. And the agentic moat I described yesterday (the seven harness files) is the operator side of the same coin. When the seven files live as engineering inside your codebase, Microsoft’s audit log only sees what you decided to show.

What changed this week is not the technology. The executive committee question is no longer which model to use. It is who collects the toll on the answer to that question, and whether you have a voice in who that someone is.