The announcement that mattered more
On June 16, 2026, Cowork reached general availability. The product news was the headline. The pricing model, buried beneath it, may prove more consequential. Cowork now runs on Copilot Credits. Each credit costs a cent, and every task draws down credits based on the model used, the depth of reasoning required, the tools invoked, and the complexity of the workflow. Individual tasks can range from a few cents to several dollars. Instead of relying on a flat license alone, Microsoft has added a consumption-based layer—not a subscription tweak—but treating AI less like software and more like cloud compute. And pricing models do more than set revenue. They shape behavior, budgeting, and governance. Every major platform reaches a point where its commercial model matters as much as its tech. Cloud computing did. Enterprise AI may arrive at that same inflection point soon.
This wasn’t a one-off decision
A few weeks earlier, GitHub Copilot made a similar move, and it’s worth pausing on why. When GitHub Copilot launched in 2022, the pricing was refreshingly simple. A flat monthly fee, unlimited use. That made sense, because Copilot mostly generated code completions and answered programming questions. Then it stopped being simple. GitHub Copilot grew into something that could review pull requests, analyze full repositories, fix bugs, and plan implementations on its own. Running an autonomous agent for thirty minutes is not the same transaction as autocompleting a line of code, yet customers paid for both identically.
So, on April 27, 2026, GitHub Copilot announced AI Credits, with usage-based billing from June 1. Simple completions stayed unlimited. But chat, autonomous agent sessions, and code review now draw down credits based on actual token usage. Seen alone, that’s a pricing change. Seen alongside Cowork, it looks like a principle. Stop charging for access to AI. Start charging for the amount of AI work performed.
Where the logic holds, and where could break
The engineering economics behind GitHub Copilot are easy to justify. An agent working across systems for half an hour consumes far more compute than a one-line autocomplete. A flat subscription covering both was never going to hold. The economics get interesting once you ask who is paying, and for what. Take a developer using GitHub Copilot. If AI ships a feature three days early, the payoff is legible. Organizations already track output through deployment frequency, cycle time, and features shipped. Faster development lands directly in numbers that leadership watches.
Now take a finance leader using Cowork. Say it assembles tomorrow’s board deck in thirty minutes instead of three hours. The productivity gain is obvious. But then what? The meeting still happens at the same time. The deck still ships once. The executive doesn’t produce five extra board decks that week. The value is real. It just shows up somewhere harder to quantify. Perhaps, the deck is better informed. Maybe, the executive spends the saved time preparing for hard questions instead of formatting slides. Maybe, it is simply one less late night. None of it lands on a spreadsheet the way a shipped feature does. Here’s some more information on what CFOs must know about AI unit economics.
The case for consumption pricing, and why it still cuts against Cowork
Buyers already accept consumption pricing for cloud, so they will accept it for AI. That’s a fair assumption. But cloud consumption maps to measurable performance. Knowledge-work value doesn’t, which is precisely the problem this model must solve. This is the crux. With developers, AI usage and business output move in near lockstep. With knowledge workers, that link loosens. Value shows up as reduced friction, and friction is notoriously hard to price. That is what makes consumption billing a bigger gamble on Cowork than it ever was on GitHub Copilot. It is also where intelligence per token becomes the number to watch. The unit economics may matter less for the cost of an answer and more for the value that answer produces. Ten dollars of AI that prevents a poor decision is extraordinarily cheap. Ten dollars spent on work an employee could have done just as well is waste. Telling the two apart is the real enterprise challenge.
