In the fourth quarter of 2026, Microsoft begins splitting Microsoft 365 Copilot into two billing tracks: a flat per-user licence for everyday AI, and metered usage-based billing in Copilot Credits for frontier models and advanced agentic work. The change, tracked in Message Center item MC1479276, marks the moment enterprise AI pricing stopped looking like a seat licence and started looking like a cloud bill. For every CIO and CFO who signed an annual Copilot contract, the AI line item is becoming variable for the first time.
The shift is not a Microsoft quirk. GitHub moved all Copilot plans to usage-based AI Credits on June 1, 2026, and Gartner now predicts that AI inference costs per agentic workflow will increase more than fivefold through 2028. The pattern is consistent across vendors: flat-rate AI is ending, and the replacement is metered consumption that rises faster than the price of tokens falls.
Microsoft splits Copilot into two billing tracks
Under the new model, "everyday AI" features such as model selection and Auto routing stay covered by the existing Copilot user subscription licence. Advanced AI and new frontier models move behind usage-based billing measured in Copilot Credits, which Microsoft describes as "a common currency for eligible Microsoft services with usage-based billing."
The critical governance detail is that nothing activates by default. Usage-based billing entry points become visible to every user, but they consume nothing until an administrator configures a spending policy in the Microsoft 365 admin center under Copilot then Cost management. Usage-based billing starts rolling out in Q4 of calendar year 2026, including as part of the Frontier public preview experience.
The same release, MC1479277, rebuilds the Copilot app around a unified Home screen plus two new destinations: Code and Autopilot, an always-on personal agent. Both are billed through usage-based billing. Copilot in SharePoint and OneDrive reach general availability, with everyday capabilities on the standard licence and new agentic, at-scale work moving to metered billing. New FinOps for AI capabilities extend cost management beyond Copilot Cowork and Work IQ APIs to include Copilot Code, Copilot Managed Runtime, and advanced SharePoint and OneDrive work, with Copilot Studio agents joining in October 2026.
Two deadlines deserve attention. End-user visibility into Copilot Credit usage became generally available in September 2026, and existing PowerShell opt-outs for the SharePoint Copilot preview are honoured only until November 1, 2026, after which administration moves to SharePoint site AI settings. A new "Request access / Learn more" admin setting reaches standard customers in October 2026 and deferred customers in November 2026, which is why Microsoft's own guidance is to set spending defaults before users see the prompts.
What a Copilot Credit actually costs
The underlying economics follow a now-familiar pattern. A Copilot Studio capacity pack lists at $200 per month for 25,000 Copilot Credits, and pay-as-you-go credits price at $0.01 each. Microsoft's credit rates, effective July 9, 2026, run from 1 credit for a classic answer up to 10 credits for tenant graph grounding, so the cost of a single interaction depends heavily on how much enterprise context the agent must pull in. Enterprises can also pre-purchase credits across volume tiers from 300,000 to 300 million credits per year, and unused credits do not roll over.
The flat-rate part of the bill has not shrunk. Microsoft 365 Copilot Enterprise still lists at $30 per user per month on an annual commitment and still requires a qualifying base licence, with Microsoft 365 E3 at $36 and E5 at $57 per user per month. That stacking is the part buyers miss: as one pricing breakdown puts it, a 100-user enterprise on E3 plus Copilot runs roughly $6,600 per month before a single Copilot Credit is consumed. The metered layer sits on top of an already substantial fixed floor.
Microsoft has not published volume discounts, though trade reporting points to possible 30 percent discounts at 1,000 to 9,999 seats and 50 percent above 10,000 seats starting as early as October 2026. Those terms are reported, not confirmed on Microsoft's public pricing pages, so they should be treated as directional until published.
GitHub Copilot already ran this experiment
Developers saw the model first. On June 1, 2026, all GitHub Copilot plans moved to usage-based billing, where one AI Credit equals $0.01 and consumption is driven by input, output, and cached tokens at each model's rates. Credits replaced premium request units. Base prices did not change: Copilot Pro stayed $10 per month, Pro+ $39, Business $19 per user per month, and Enterprise $39.
What changed was what each plan includes. Copilot Business includes 1,900 credits per user per month and Enterprise includes 3,900, both pooled at the billing entity. GitHub softened the transition with promotional allotments from June 1 through September 1, 2026 of 3,000 credits per Business seat and 7,000 per Enterprise seat. When the promotion ended on September 1, Enterprise pools shrank 44 percent and Business pools shrank 37 percent, with no change to the seat price and no change to how engineers work. Code completions and Next Edit suggestions remained included and do not consume credits, but everything more ambitious does.
The GitHub rollout is the cleanest preview of what enterprises should expect from the Microsoft rollout: a fixed seat that no longer caps the bill, a pooled credit balance that quietly tightens, and a governance surface that finance teams must now operate.
The Inference Paradox: cheaper tokens, bigger bills
The reason usage-based AI billing keeps surprising buyers is that unit prices and total spend are moving in opposite directions. Gartner's August 17, 2026 press release predicts that AI inference costs per agentic workflow will increase more than fivefold through 2028, an effect it names the Inference Paradox.
Gartner identifies three forces. Foundational model cost economics keep improving. That efficiency unlocks the deployment of more powerful, and more expensive, models. And those more sophisticated workflows use far more tokens than simple chatbot interactions. As Will Sommer, Senior Director Analyst at Gartner, put it: "Product leaders cannot rely on more efficient token economics to rationalize AI costs. Each successive generation of AI capability will necessitate more, and often more expensive, tokens. There is no reliable, economical one-size-fits-all model on the horizon."
The mechanics are stark. Gartner states that routing a task to an agentic reasoning model increases provider inference costs by at least five times compared with a basic chatbot, and often much more as task complexity grows. Sommer warned that "defaulting to generic autonomous intelligence will result in unbounded costs orders of magnitude higher than those of optimized product ecosystems."
Why agentic workflows break the flat-rate budget
The gap between falling token prices and rising bills shows up in enterprise budgets. The FinOps Foundation's 2026 State of FinOps report found that 73 percent of enterprises reported AI costs exceeding original projections, even as token prices fell roughly 98 percent between 2022 and 2026. The same analysis found a tiered model architecture achieved a median blended cost of about $2.31 per million tokens, while routing every workload to frontier models cost about $18.40 per million, an 87 percent gap driven by one architectural decision.
The individual agent interaction tells the same story. EY's analysis found the cost of a single agentic customer-service interaction rose from about $0.04 in 2023 to $1.20 in 2026, roughly a 30-fold increase, because orchestrated multi-tool workflows replaced simple linear exchanges. Gartner data cited by the same research shows agentic workflows burn 5 to 30 times more tokens per task than a basic query.
The corporate consequence is already visible. A Gartner poll found that more than 40 percent of agentic AI projects will be canceled by the end of 2027 due to escalating costs, unclear business value, or inadequate risk controls. The billing model and the project cancellation rate are two views of the same problem: AI value is not scaling as fast as AI consumption.
What enterprises should do now
The Microsoft rollout gives administrators a narrow window, and the recommended actions are concrete. Configure spending policies in the Microsoft 365 admin center under Copilot then Cost management before users start hitting usage-based billing prompts. Confirm that cost management policies extend to the newly metered SharePoint, OneDrive, and Copilot Studio agent capabilities. Decide the default for the new access setting before the October and November rollout reaches your tenant.
Beyond configuration, the durable levers are architectural. Gartner's advice is to calibrate complex tasks against cheaper intelligence through inference tiering, routing, and orchestration rather than defaulting every request to a frontier model. Practical steps include prompt caching, smart model routing, agentic spend controls, and treating AI-assisted work with the same visibility and governance that teams already apply to cloud and CI/CD capacity. The lesson from both GitHub and Microsoft is that pooled credits create a shared budget that can be exhausted by a few heavy users, so per-user and per-agent consumption reporting is now a finance function, not just an IT one.
What to watch next
Three signals will show whether this transition is managed or chaotic. First, whether Microsoft publishes the volume discount tiers it has been reported to be preparing, since those terms determine whether large buyers absorb the shift or renegotiate. Second, whether the GitHub pooled-credit step-down of September 2026 triggers visible pullback in engineering usage, which would be an early read on how enterprises respond when a free allowance tightens. Third, whether the Inference Paradox shows up as a wave of agent project cancellations in 2027 or as a discipline of tiered, measured deployment.
The through-line is simple. AI is no longer sold as a fixed feature with a fixed price. It is sold as metered intelligence, and the invoice now reflects how aggressively a company chooses to use it. The organizations that treat AI consumption like cloud spend, with real architecture, real limits, and real FinOps, will be the ones that keep the productivity gains without discovering that their flat-rate assumptions quietly expired.