"We approved Copilot and Claude as a fixed per-seat cost. Why is finance now asking me to forecast tokens?" We hear some version of that question from operations leaders every renewal season now, and it is a fair one. GitHub Copilot, Claude Enterprise, and parts of Microsoft 365 Copilot have all moved toward usage-based billing, which means the line item you signed off on last year no longer behaves like a subscription.
The short answer for anyone skimming: AI usage-based billing replaces a predictable seat fee with metered consumption, so the real cost now depends on which models people use and how much agentic work runs unattended. The controls to cap that spend exist on every major platform, but several of them are off or unlimited by default. Someone has to switch them on and own them.
GitHub moved every Copilot plan to usage-based billing on June 1, 2026. Under the old model, each interaction cost a premium request with a model multiplier. Under the new one, GitHub's announcement explains that premium requests were replaced by GitHub AI Credits, consumed according to input, output, and cached tokens at each model's published API rate. Base prices did not move: Business stays at $19 per user per month and Enterprise at $39, with each seat including AI Credits worth the same amount. Code completions and next edit suggestions remain included and do not consume credits.
Two details in that announcement matter more than the headline. First, the old fallback is gone. Users who exhausted their premium requests used to drop to a cheaper model and keep working; now usage is governed by available credits and admin budget controls. Second, Copilot code review began consuming GitHub Actions minutes in addition to AI Credits, which the June 1 GitHub changelog confirmed alongside the general availability of user-level budgets. A tool your developers think of as "the Copilot license" now touches two separate meters.
Anthropic's change is structural rather than incremental. According to Anthropic's Enterprise billing documentation, the current Enterprise plan has two parts: a seat fee that covers access to Claude, Claude Code, and Cowork, and separate usage charges for every token at standard API rates. There is no included token allowance per seat. Organizations on the older Standard and Premium seat plans keep their terms until their next contract renewal, when they transition automatically to the single usage-based Enterprise seat.
That renewal clause is where budgets get blindsided. A company that bought Premium seats with a generous built-in allowance can renew into a model where the seat fee buys access and everything else is metered. Nothing about how employees use Claude has to change for the invoice to look completely different.
Anthropic's handling of its Agent SDK is the other lesson. It announced in May that automated usage through the Agent SDK and the claude -p command would move to a separate monthly credit, then, as its support article on the Agent SDK records, paused that change on June 15 before it took effect. On October 7 it added monthly API credits to Max and Team plans for API-key usage. Three announcements in five months is not a criticism of Anthropic. It is the pace your budget process now has to absorb.
The common belief is that usage-based billing is a disguised price increase. For most small and mid-sized teams, it is not. GitHub's per-seat prices are unchanged, and many users will never touch the bottom of their included credits. The risk is that the controls protecting you from a runaway month are not configured for you, and in some cases are explicitly off.
GitHub's own documentation on Copilot budgets is blunt about this. The "Stop usage when budget limit is reached" setting for enterprise, cost center, and organization budgets is off by default, and without it charges continue to accrue past the limit. The same page warns that an enterprise budget is not a total monthly budget; it only caps metered charges after the shared credit pool is exhausted. GitHub's worked example puts 400 Business licenses at $7,600 in license fees plus a $5,000 enterprise budget, for a maximum bill of $12,600. If nobody reads that page, the budget you set feels like a ceiling and behaves like a suggestion.
Anthropic lets owners set organization and individual spend limits, but its billing documentation also notes that owners can set those limits to "unlimited" and consumption will still be billed. Microsoft takes a similar approach with Copilot Credits: its guidance on managing spending policies for usage-based Copilot services shows policy scope defaulting to all users and an "Auto-apply new services" setting that is on by default, so newly supported agents inherit your policy and its billing method automatically. That is convenient, and it also means new metered services can start drawing credits without a fresh decision from anyone.
| Platform | What you are billed on | What happens at the limit | Default worth checking |
|---|---|---|---|
| GitHub Copilot Business and Enterprise | Seat fee plus pooled AI Credits; metered at $0.01 per credit after the pool | User budgets are hard stops; other budgets stop usage only if configured to | "Stop usage when budget limit is reached" is off by default |
| Claude Enterprise (usage-based) | Seat fee for access; all tokens at standard API rates | Spend limits stop usage; self-serve stops when prepaid credits run out | Limits can be set to unlimited and still billed |
| Microsoft 365 Copilot usage-based services | Copilot Credits through prepaid packs, pre-purchase, or pay-as-you-go | Users lose access to covered agents until credits reset | "Auto-apply new services" is on by default |
Agentic workflows change the math because cost now scales with work performed, not with the number of people logged in. A developer accepting completions barely registers. The same developer asking an agent to refactor ten files, run tests, and iterate will consume tokens on every loop, and none of those loops wait for a human to approve the spend.
This is why the distinction GitHub and Anthropic both draw between interactive use and automated use matters to a business leader, not only to engineering. Automated code review, scheduled agents, and pipeline integrations are the usage most likely to grow quietly, and they are the hardest to attribute to a specific person when the invoice arrives. If you have read our piece on who is supervising your AI agents, this is the financial version of the same supervision problem.
Treat AI spend the way a well-run company treats cloud spend: owned, capped, reviewed, and attributed. The steps below are the ones we see close the gap fastest in 20 to 300 person organizations, and none of them requires new tooling.
That last step is where most teams stall, because nobody agreed upfront on what the AI was supposed to produce. Our guide to proving whether your AI agents are paying off covers how to set that baseline, and the broader question of what AI actually costs a small business is worth revisiting now that consumption, not seats, drives the number.
It can be, if you manage it. A flat seat fee hid the cost of heavy users inside the cost of light ones, which made it impossible to tell whether a tool was earning its keep. Metered billing exposes exactly who and what is consuming capacity, and that visibility is useful the moment you start asking whether a workflow is worth its tokens.
The catch is that visibility only helps if someone looks. Companies that already run disciplined technology planning and budgeting will absorb this shift as one more metered service. Companies that approved AI tools as a one-time decision, without an owner, are the ones that will see a surprise and conclude AI "got expensive." In our experience, that reaction is also how many SMB AI initiatives stall: not because the tool failed, but because nobody could explain the bill.
It also strengthens the case for keeping your options open. When each vendor can change its billing model at renewal, your ability to move workloads matters. We cover that in our piece on AI vendor exit plans, and the same logic applies to pricing as it does to outages.
When we work with a client on this, we start with an inventory of every AI platform with a billing relationship, who holds admin rights in each, and which budget and limit settings are currently active. In most environments we find at least one hard stop that was never enabled and at least one automated workflow drawing from the same pool as everyone else. We then configure the caps, split automated usage into its own budget, and set up the monthly review so finance sees AI consumption next to the outcomes it is supposed to support.
From there, the work is ongoing rather than one-time. Vendors will keep adjusting their models, and our role is to track those changes against your renewal calendar so the next billing shift is a planned conversation with finance rather than an unwelcome surprise.
If your AI tools are billed by usage but still budgeted like subscriptions, the fix starts with ownership and caps, not with cutting access. Make sure the people using these tools understand what drives cost, especially when they launch agents that run without supervision.
If your team is moving faster with AI than your guardrails are, start with structured training rather than another tool. AI University gives your people role-based AI training with security built into the material, not bolted on afterward.
If you would rather talk through your specific environment first, book a strategy call with Securafy and we will walk your current AI usage, exposure, and the fastest path to safe adoption.