Microsoft just told its biggest AI customers something they were not expecting: the more you use Copilot Cowork, the more you will pay, and the cheap tier might run on a Chinese AI model that everyone is still arguing about.
On June 16, 2026, Microsoft launched Copilot Cowork globally with usage based pricing, dropping the flat rate model that defined Microsoft’s AI products for the past two years. The same week, Axios reported that Microsoft is testing a self hosted, fine tuned version of DeepSeek V4 as a low cost alternative to the Anthropic and OpenAI models that currently power the tool. Two separate decisions, one underlying problem. Agentic AI is expensive to run, and Microsoft just admitted it out loud.
What Actually Changed
Copilot Cowork is Microsoft’s agentic assistant inside the Microsoft 365 ecosystem. Unlike a simple chat prompt, it can carry out multi step tasks across Outlook, Teams, and Excel on its own, checking its own work and calling the model again and again until the job is done.
That loop is exactly what makes it expensive. Every time the agent rereads a document, rewrites a draft, or double checks an answer, it burns more tokens. Under flat rate pricing, Microsoft was absorbing that cost. Starting now, customers pay based on how much compute their Cowork agents actually consume.
Companies using Cowork can currently choose between Anthropic’s Opus 4.8 and Sonnet 4.6, while customers enrolled in Microsoft’s Frontier program get access to OpenAI’s GPT 5.5 and an in house model called Cowork 1, according to Computerworld’s reporting on the launch. None of those are cheap to run at scale. That is where DeepSeek comes in.
Why Microsoft Is Doing This Now
Charles Lamanna, Microsoft’s executive vice president for Copilot, agents and platform, gave Axios the blunt version of the math. Power users running hundreds of agent tasks a week are exactly the customers Microsoft wants, and exactly the customers an unlimited plan cannot survive. As Lamanna put it, those users are “way productive,” but past a certain point, “the costs can go very high.”
That tension is not unique to Microsoft. Every company building agentic tools is running into the same wall. Anthropic’s Claude Code and OpenAI’s Codex work the same way Cowork does: an agent that keeps calling the model on its own as it works through a task, instead of waiting for a human to type the next prompt. More autonomy means more token consumption, and more token consumption means the old subscription math stops adding up.
Microsoft already ran this experiment once. GitHub Copilot switched from flat premium request units to token metered GitHub AI Credits on June 1, 2026, where one credit equals one cent and usage is calculated from input, output, and cached tokens at each model’s published rate. Cowork’s shift to usage based billing follows the same logic, just two weeks later and at the level of an entire enterprise productivity suite.
The DeepSeek Question
Here is the part that will actually generate headlines. DeepSeek is a Chinese AI lab known for building open source models that perform close to top tier Western systems at a fraction of the inference cost. Microsoft telling Axios that it is exploring a fine tuned, Azure hosted version of DeepSeek V4 for an enterprise product as mainstream as Copilot Cowork is a different conversation than quietly offering the same models as a developer option, which Microsoft has done through Azure AI Foundry since January 2025, as TheStreet noted in its coverage of the pricing shift.
Microsoft is aware of how this looks. Axios flagged plainly that adding a model from a Chinese AI company could draw criticism, especially in the US, given the ongoing scrutiny of Chinese AI labs and where enterprise data ends up. Microsoft’s answer, according to its statement to Axios, comes with three conditions attached:
- Optional, not default. Customers would choose whether to use the DeepSeek backed tier at all.
- Fully hosted on Azure. Customer data stays inside Microsoft’s cloud, covered by Azure’s existing enterprise security, compliance, and data residency controls rather than touching any DeepSeek owned infrastructure.
- Fine tuned with added safeguards. Microsoft says it has modified the model and built in changes aimed at reducing bias before considering it for production use.
Microsoft has not confirmed it will actually ship DeepSeek. The company told Axios it expects to make a lower cost model option available in the coming weeks and will confirm its final choice then, which could mean DeepSeek V4, a different open source model, or something else entirely.
The price gap that’s driving this decision is significant. One Chinese tech outlet’s reporting, cited by Let’s Data Science, put Anthropic’s flagship model at roughly $50 per million tokens against DeepSeek V4 Pro at roughly $0.87 per million tokens. Treat that exact figure as a rough estimate rather than a confirmed rate sheet, since it comes from a single secondary source. But even directionally, it explains why Microsoft is willing to absorb the political risk of putting a Chinese model inside an American enterprise product.
| Factor | Anthropic / OpenAI models (current default) | DeepSeek V4 (proposed cheap tier) |
|---|---|---|
| Reported relative cost | Premium pricing, reported around $50 per million tokens for the flagship Anthropic model | Reportedly a small fraction of that, around $0.87 per million tokens for DeepSeek V4 Pro |
| Hosting | Provided through existing Anthropic and OpenAI agreements | Would run fully on Azure under Microsoft’s own infrastructure |
| Customer choice | Default option today | Proposed as optional, not a forced replacement |
| Status | Live now in Copilot Cowork | Still being tested, no confirmed launch date |
The Bigger Pattern Behind Both Moves
Step back and the two announcements are really one story. Microsoft is building toward a multi model approach instead of betting everything on a single AI partner. That gives the company room to offer a premium tier for complex reasoning and a budget tier for routine, repetitive agent work, without eating the cost difference itself.
It also signals something about where the whole agentic AI category is headed. The era of paying one flat fee and using an AI agent as much as you want is closing across the industry, not just at Microsoft. If your team relies on any tool that lets an AI agent work autonomously through multi step tasks, the billing model under it is probably changing too, whether or not the vendor has announced it yet.
What To Do If You Use Copilot Cowork, or Any Agentic AI Tool
You do not need to wait for a confirmed DeepSeek decision to get ahead of this. The same steps protect you regardless of which model ends up running your cheap tier.
- Audit last month’s usage before the new pricing kicks in fully. Pull a report of how many agent tasks your team actually ran and which ones took the most back and forth. That is your baseline for what the new bill will look like.
- Set hard budget caps, not just alerts. A warning email after you’ve already burned through the budget does not help. Look for a setting that stops or pauses usage once a spending limit is hit, the same way GitHub lets Copilot users cap their AI Credits.
- Match the model to the task, not the other way around. Simple, repetitive work like formatting a spreadsheet or drafting a routine email does not need your most expensive reasoning model. Save the premium tier for genuinely complex, multi step problems.
- Ask your vendor for a cost preview before you commit to a workflow. If Microsoft or any other provider offers a projected bill or usage dashboard, use it before scaling an agent workflow across your whole team, not after.
- Test any new cheap model on real work before trusting it at scale. If Microsoft does roll out a DeepSeek backed tier, run your actual tasks through it first. Compare accuracy and output quality against your current model before moving anything business critical onto it.
- Avoid building workflows that depend entirely on one vendor’s pricing decisions. Keep at least one alternative tool or model in your back pocket. The companies offering flat rate AI access today are the ones most likely to follow Microsoft’s lead tomorrow.
The Honest Caveats
None of this is final yet. Microsoft has not confirmed which model will actually power the lower cost tier, and “the coming weeks” is the only timeline the company has given. The exact compute based pricing structure for Copilot Cowork, meaning the actual rate per task or per token customers will pay, has not been published in full detail either.
The DeepSeek question also is not purely technical. Even with Azure hosting, fine tuning, and bias safeguards in place, some regulated industries and government linked organizations may avoid any model with origins in a Chinese AI lab regardless of where it is hosted, simply on policy grounds. That is a compliance conversation your legal or procurement team needs to have separately from the cost conversation.
And the cost comparison numbers floating around right now, including the one in the table above, come from secondary reporting rather than an official Microsoft or DeepSeek rate card. Treat them as directionally useful, not as numbers to budget against.
The flat rate AI subscription was always a bet that usage would stay predictable. Agentic AI just proved that bet wrong, and now everyone from Microsoft down to your own team has to figure out what AI actually costs to use well.
Sources
- Axios: Microsoft weighs DeepSeek for Copilot Cowork (June 16, 2026)
- Computerworld: Microsoft launches Copilot Cowork with usage based pricing
- The GitHub Blog: GitHub Copilot is moving to usage based billing
- Let’s Data Science: Microsoft explores DeepSeek V4 for Copilot Cowork costs
- TheStreet: Microsoft changes Copilot pricing model
Just Being Resourceful! 
Leave a comment