Picture someone at a company who has fully bought into Microsoft Copilot. They use it constantly, hundreds of tasks a week, drafting documents, summarizing meetings, chasing down information across the company’s files. By any measure, exactly the kind of power user Microsoft would want.
Turns out that person is also expensive to serve.
“We have users who do hundreds of tasks a week, which is great , they’re way productive , but the consequence is the costs can go very high,” Charles Lamanna, Microsoft’s executive vice president for Copilot, told Axios on June 16. Lamanna runs Copilot. And he was essentially telling a reporter that the way the product had been priced could not hold.
What Microsoft is changing about how Copilot charges customers
The announcement that came alongside that quote was that Copilot Cowork, Microsoft’s enterprise AI agent product, is moving to usage-based pricing. Instead of a flat monthly rate per user, companies will pay based on how much compute they actually use, through what Microsoft is calling Copilot Credits.
Here is the thing about agentic tools. A chatbot answers a question and stops. An agent keeps going, reading files, calling other tools, checking its own work, trying again, burning compute at every step.
Related: Microsoft and Copilot just hit a jackpot in healthcare
A user running an agent hundreds of times a week consumes a completely different amount of compute than someone who asks one question and closes the tab.
Flat pricing was never going to survive that reality at scale. Microsoft learned this the hard way with GitHub Copilot, which moved to token-based billing on June 1, generating significant developer backlash after some users saw monthly costs balloon from $29 to nearly $750, TechCrunch reported.
Microsoft, to its credit, is not pretending otherwise. Lamanna’s quote is unusually blunt for a product executive. It confirms what anyone building or buying agentic AI has probably suspected: the bills get real very fast once people actually start using these tools every day.
Where DeepSeek fits into Microsoft’s plan for Copilot
The second piece of the announcement is the one getting more attention. Microsoft told Axios it is exploring a fine-tuned version of DeepSeek V4, or possibly another open-source model, as a cheaper alternative to the Anthropic and OpenAI models currently powering Copilot Cowork. No final decision has been made, and Microsoft says it will confirm its choice in the coming weeks.
DeepSeek is a Chinese AI lab whose models deliver strong benchmark performance at much lower inference costs than the big frontier models. That cost gap matters a lot for an agentic product, where every workflow chains together dozens of model calls. Microsoft is not abandoning OpenAI or Anthropic for complex tasks. The thinking seems to be tiered: expensive models when genuinely needed, something cheaper when they are not.
More Tech:
- Microsoft CEO sends a blunt warning on AI and the tech ecosystem
- Amazon CEO just made things uncomfortable for Anthropic
- Microsoft has bad news for a key AI partner
The catch, and it is a real one, is that DeepSeek is Chinese. Axios flagged that the move could draw criticism. Microsoft’s answer is that if DeepSeek is used, it would run entirely on Azure, under the same enterprise security and compliance controls that govern everything else on the platform.
DeepSeek’s models have actually been available through Azure AI Foundry since January 2025, so this is not the first time Microsoft has put them inside its infrastructure. Building them into a mainstream enterprise product like Copilot Cowork is a different conversation than offering them as a developer option, though. The GitHub Blog framed the broader shift explicitly: “agentic usage is becoming the default,” driving higher compute and inference demands that flat-rate models were not built to absorb.
Sugita/Getty Images
What this means for Microsoft stock investors
The pricing change and the DeepSeek exploration are connected. They are both about the same underlying problem: how do you make an agentic AI product that people love to use without the economics falling apart? Flat-rate pricing does not work, and expensive models on every step do not work either.
For investors, the usage-based model is actually the more significant shift of the two. Moving from seat licenses to consumption billing changes how Microsoft reports and forecasts AI revenue, and introduces variability that analysts have not had to model in Microsoft’s enterprise software business before. That is not necessarily bad, but it will take a few quarters before anyone knows what a sustainable usage average looks like at scale.
The DeepSeek side still has an open question attached to it. Microsoft has not confirmed which model it will actually use, and investors should probably wait for that announcement before drawing firm conclusions about what it means for margins or for the OpenAI relationship.
What is already clear is that Microsoft is serious about making Copilot work as a business, not just as a product demo. Sometimes that means saying out loud that the current setup has a problem.
