On June 1, 2026, Microsoft’s GitHub division changed how it charges for Copilot. The announcement, made April 27 by GitHub VP Mario Rodriguez, framed it as a technical billing update: Premium Request Units — the fixed metering system that had governed Copilot charges since the product’s enterprise launch — were being replaced by GitHub AI Credits, where one credit equals $0.01 and consumption reflects actual token usage at published API rates. Base subscription prices, Rodriguez noted, remain nominally unchanged.
That last sentence is doing a lot of work. For developers whose Copilot use consists of autocomplete and single-turn code suggestions, the change is close to neutral. For developers running agentic workflows — multi-step, multi-model tasks where an AI agent breaks a prompt into subproblems, executes each with a different model, synthesizes the output, and loops until done — token consumption per session is an order of magnitude higher than a standard code suggestion. For those users, the bill just got structurally larger, and a three-month promotional credit ($30 per Business seat, $70 per Enterprise seat through August 2026) is the only buffer between the old cost structure and the new one.
The billing change is not an isolated product update. It is the most visible sign yet of a pressure Microsoft has been managing for two years: the compute economics of AI are straining the pricing models that were designed to sell AI as a fixed-cost productivity layer, and the company that invested approximately $190 billion in AI infrastructure for 2026 alone needs its AI products to generate returns that justify that number. GitHub Copilot’s shift to token billing is how Microsoft begins moving the cost of agentic AI from its own balance sheet onto its customers’.
What the Billing Change Actually Means
Under the old system, GitHub Copilot plans came with a defined monthly allocation of Premium Request Units. When those units were exhausted, users could either stop using premium model features for the rest of the month or purchase additional units. The model was predictable: a team could budget Copilot costs with the same certainty as a SaaS license, because the ceiling was fixed.
Under the new system, there is no fixed ceiling for standard token consumption. GitHub AI Credits are debited as requests are processed, at rates that reflect the actual compute cost of each model call. A request routed to a lightweight model costs fewer credits than a request routed to a frontier reasoning model. An agentic workflow that chains five model calls in a single session costs five times more than a single-call interaction, plus any additional tokens generated by the agent’s internal reasoning steps. GitHub’s published June 1 changelog entry makes this explicit: billing now reflects ‘actual token consumption at published API rates.’
The rate limit changes that accompanied the billing update are equally significant. Across Copilot Business, Copilot Enterprise, and individual plans, GitHub tightened the monthly caps on premium model requests. The practical effect for heavy agentic users is that the old soft limit — burn through your PRUs and the product continues working at a degraded model tier — has become a harder cost boundary, where additional usage accrues charges rather than degrading gracefully to a cheaper model.
The promotional credits bridge the transition. Copilot Business customers receive $30 per user per month in GitHub AI Credits for June, July, and August 2026. Enterprise customers receive $70. At the $0.01 per credit rate, Business users get 3,000 credits per month and Enterprise users get 7,000. What the promotional period absorbs in actual usage before the credits are exhausted depends entirely on the model mix and session depth of each developer’s workflow. GitHub has not published conversion rates from the old PRU system to the new credit system, making the direct cost comparison between old and new difficult to calculate precisely in advance — which is itself a source of enterprise frustration with the change.
Copilot Studio Did This Nine Months Earlier
The GitHub Copilot change is not the first time Microsoft has moved a Copilot product away from a fixed-unit billing model toward token-based consumption. Microsoft Copilot Studio — the low-code platform for building custom AI agents on Microsoft’s infrastructure — made an analogous shift on September 1, 2025, when it rebranded its billing unit from ‘messages’ to ‘Copilot Credits.’
The Copilot Studio transition was structured to appear change-neutral: the prepaid capacity pack remained 25,000 credits per month for $200 per month, and Microsoft’s documentation explicitly stated that ‘there’s no change in the quantity per prepaid pack or to the pay-as-you-go rate.’ The pay-as-you-go rate for Copilot Studio through Azure subscription is $0.01 per credit — the same unit price that GitHub is now applying to Copilot developer billing. The underlying billing architecture is identical across both products.
The September 2025 Copilot Studio change was framed at the time as an administrative simplification — a common currency across Microsoft’s agent platform. In retrospect, it established the pricing infrastructure that GitHub’s June 2026 change builds on. Microsoft has been migrating its AI products toward a token-consumption billing standard since at least mid-2025, with each individual product change framed as a standalone update rather than an acknowledged architectural shift. The aggregate effect is a Microsoft AI portfolio where variable, consumption-based costs are replacing predictable fixed fees across the stack.
The Agentic Compute Problem Microsoft Is Solving For
GitHub’s own announcement language explains the economic pressure driving the change. ‘Agentic usage is becoming the default,’ Rodriguez wrote in the April 27 blog post, ‘and it brings significantly higher compute and inference demands.’ The shift from AI-as-autocomplete to AI-as-agent is not a marginal increase in resource consumption. A single agentic coding session — where a developer prompts Copilot to implement a feature, tests and debugs iteratively, generates documentation, and writes test cases — can consume token volumes that dwarf a month of traditional autocomplete interactions.
The compute economics of agentic AI are structurally different from the compute economics of the single-turn AI that dominated 2023 and 2024. In a single-turn model, a user sends a prompt and receives a response; the infrastructure cost is bounded by the length of the prompt plus the response. In an agentic model, an AI system plans, executes multiple steps autonomously, evaluates intermediate outputs, and iterates — each step generating its own prompt-response cycle, often routed through frontier models that cost substantially more per token than the base models used for simple completions. GitHub has actively promoted these agentic capabilities: Copilot Workspace, agent mode in Visual Studio Code, and the Copilot Extensions platform were all shipped and marketed through 2024 and 2025 specifically to drive agentic adoption at depth. The billing model was not updated to reflect the resulting economics until June 2026.
Microsoft’s promotional credit amounts offer an implicit reveal of how much heavy agentic consumption costs the company per seat. Enterprise customers receive $70 per month in credits — 7,000 credits at $0.01 each. If those credits represent a reasonable consumption budget for typical Enterprise usage, the implied monthly compute cost per active Enterprise seat is somewhere in the $70 to $150 range, depending on model mix and session depth. Against an Enterprise subscription price of approximately $39 per user per month, that implies a range of scenarios where Microsoft’s cost of serving a heavy agentic Enterprise user exceeds the subscription revenue from that user. The promotional credit amount is calibrated to cover what Microsoft expects the average heavy user to consume — which is why Enterprise gets more than double the Business allocation. After August, those costs transfer to the customer.
Rodriguez’s statement that the change is ‘an important step toward a sustainable, reliable Copilot business’ is as direct an acknowledgment as a product announcement typically offers that the old model was not sustainable at the consumption levels agentic usage generates. The transition to token billing ensures that as usage intensity increases — which Microsoft’s own feature roadmap is designed to drive — revenue scales with it rather than running at an ever-widening deficit against infrastructure costs.
The $190 Billion Problem Behind the Billing Change
The compute sustainability argument for usage-based billing is real and would exist regardless of Microsoft’s financial position. But it does not exist in isolation from that position. Microsoft CFO Amy Hood disclosed during the company’s Q3 FY2026 earnings call on April 29, 2026, that Microsoft’s capital expenditure for the full calendar year 2026 would reach approximately $190 billion — a 61 percent increase over 2025’s approximately $118 billion, and more than three times the 2024 figure. The infrastructure investment is entirely oriented toward AI: data centers, networking, and the GPU clusters required to run inference at the scale Microsoft’s AI product ambitions require.
Against that investment, Microsoft’s AI products need to generate a return that eventually justifies the $190 billion outlay. The problem is that the primary vehicle for that return — Microsoft 365 Copilot, the enterprise AI assistant bundled with the commercial M365 suite — has reached approximately 3.3 percent of its addressable market. At Microsoft’s Q2 FY2026 earnings call in January 2026, the company disclosed 15 million paid Copilot seats against a commercial M365 base of more than 450 million. By April 2026, the seat count had grown to 20 million — still under 4.5 percent penetration. The trajectory is positive; the gap to a number that justifies the capital commitment is substantial.
Microsoft’s platform position creates leverage for this kind of monetization shift that a standalone AI tool provider could not exercise. GitHub Copilot is embedded in developer workflows that have switching costs — project history, institutional knowledge of the tool, integration with GitHub Actions and pull request workflows. That embeddedness gives Microsoft the ability to change pricing terms in ways that a product without those switching costs could not. The usage-based shift is, in part, an exercise of that embedded position: the product is valuable enough that most enterprise customers will absorb the billing change rather than migrate to alternatives.
The capex math makes the timing of the billing change legible. Microsoft committed capital at a scale that requires its AI products to perform well beyond their current penetration rates. Usage-based billing accelerates per-seat revenue extraction from the customers already in the product without requiring new seat sales. A Copilot Business customer who moves from simple code completion to agentic workflows — which Microsoft’s own product roadmap actively encourages — pays more per month without any sales motion required. The revenue scales with usage, and Microsoft’s incentive is to drive usage up.
What Enterprises Are Dealing With
The enterprise reaction to the billing change reflects a broader tension in the Copilot deployment story. The adoption gap — 3.3 percent penetration despite two years of aggressive Microsoft marketing — is not primarily a pricing problem. It is a job-fit problem: enterprises have struggled to identify the specific high-value workflows where Copilot demonstrably improves output, and without that identification, broad seat deployment does not produce the ROI numbers that justify expansion.
Usage-based billing adds a new dimension to that challenge. Under fixed subscription pricing, the cost of a Copilot deployment is known in advance: seats times price. A CFO approving 500 Copilot Business seats knows the monthly commitment is $9,500. Under usage-based billing, that number becomes variable — potentially higher if agentic adoption accelerates, potentially the same if teams use the product for simple completions, unpredictable in either case without close monitoring of token consumption per team and per workflow type. For enterprises already struggling to quantify AI productivity gains, unpredictable cost is an additional friction that slows expansion decisions.
Microsoft’s internal assessment of Copilot adoption — the Code Red framing that emerged from its own usage data — acknowledged that Copilot had not achieved the workflow integration depth that would produce strong retention and expansion economics. The June billing change arrives at a moment when enterprise customers are still making those workflow integration decisions. Variable billing shifts the economic risk of low-utilization deployments from Microsoft to customers: if a team pays for seats and doesn’t use them, Microsoft absorbs no additional cost; if a team uses seats heavily for agentic work, Microsoft now captures that usage economically rather than eating the compute cost against a fixed subscription price.
The variable billing structure also creates a competitive opening for AI coding alternatives that maintain flat-fee pricing. Cursor, Windsurf, and other IDE-first AI coding tools have built their enterprise growth partly on predictable subscription economics that make budget approval straightforward. If GitHub Copilot’s September billing materialized significantly above the current subscription cost for a developer cohort, those alternatives become an easier internal sell for teams that want cost certainty over model breadth. Microsoft’s embedded position in the GitHub and Azure ecosystem is the primary barrier against that substitution — but the June billing change makes the switching cost calculation more explicit for every enterprise Copilot buyer.
For enterprise IT and procurement teams, the practical response is instrumentation. The three-month promotional window through August 2026 is, in effect, a measurement period: organizations that use it to understand their actual per-developer, per-workflow token consumption will be better positioned to forecast September costs accurately. GitHub’s billing dashboard exposes credit consumption at the organization and team level. The discipline to use it before the promotional period expires is the most direct thing enterprise Copilot buyers can do to avoid a billing surprise in September.
The Question the Promotional Credits Don’t Answer
The broader question the GitHub Copilot change raises is whether it marks the beginning of a Microsoft-wide pricing model shift. Copilot Studio moved to credits in September 2025. GitHub Copilot moved to credits in June 2026. Microsoft 365 Copilot — the flagship enterprise product, with 15 to 20 million seats and a $30 per user per month price point — remains on fixed subscription billing. If the agentic compute economics that drove the GitHub change apply equally to the M365 Copilot product, and they do, the fixed M365 pricing faces the same sustainability pressure. A move toward usage-based M365 billing would be a larger and more consequential change than the GitHub update — affecting enterprise agreements across thousands of organizations that have committed to fixed-cost AI budget lines.
Microsoft has not announced any changes to M365 Copilot subscription pricing. The June 2026 GitHub update and the September 2025 Copilot Studio update are, so far, limited to the developer and agent-building products. But the direction is visible: Microsoft is moving toward a billing architecture where the cost of AI consumption is borne proportionally by the customers generating that consumption, rather than pooled across a subscriber base at a fixed price. The promotional credits buy time through August. What enterprises do with that time — instrument their usage, identify high-value agentic workflows, or defer the hard deployment decisions until billing forces clarity — will determine whether September 2026 marks a managed transition or a budget shock. And if M365 Copilot follows the same path, the September deadline becomes a preview of a much larger renegotiation between Microsoft and its enterprise customer base.
