What the Bill Can't Measure

On June 1, GitHub switched Copilot to usage-based billing. Code completions remain free. Chat, agent mode, code review — anything consuming meaningful tokens — now runs on metered credits. A single agentic coding session, where Copilot autonomously reads a codebase, plans changes across files, runs tests, and iterates, can consume $30 to $40. Under the old flat rate, that session and a one-line chat question both counted as one request. The economics pointed one direction, and GitHub followed.

The new setup is genuinely useful. Under flat-rate billing, organizations had no signal at all about where their AI investment was going. Now they get pooled credits, hierarchical budget controls, per-model token dashboards. An admin can cap a user's spending and halt access when it's exhausted. GitHub even gave organizations their April usage data in advance so they could model the new bill. When the meter runs, people pay attention.

The natural parallel is cloud infrastructure. A decade ago, organizations moved from flat-rate hosting to metered cloud and discovered they were wasting 25 to 35 percent of spend on idle resources and over-provisioned instances. FinOps emerged to find and eliminate that waste, and mature programs cut it roughly in half.

That parallel worked because the underlying measurement was tractable. An idle VM is idle. An over-provisioned instance has measurable excess capacity. "Correct" infrastructure usage is definable as a utilization target, and deviation from it shows up in the bill. Use less while maintaining performance, and performance is measurable.

AI tool output resists that kind of measurement. The FinOps Foundation's 2026 survey surfaces the gap directly. One practitioner put it plainly:

The industry shifted from "what did it cost?" to "did it produce value?" and the tooling to answer the second question doesn't exist at anything like the same maturity.

The developer-side evidence is consistent. 96% of developers don't fully trust AI-generated code, but only about half always verify it before committing. AI-generated code shows 1.7x more issues in pull requests, with logic and correctness problems 75% more common. The characteristic failure mode is code that compiles cleanly, looks plausible, and breaks in ways that surface later. Systematic misinterpretations, the same confident wrong assumption applied consistently. On the bill, they look exactly like correct output. They don't show up as retries or excess token consumption. They show up as production incidents, weeks later.

This is where the cloud analogy loses its footing. Cloud waste correlates with cost: over-provisioned instances are visible, and cutting them saves money without degrading performance. The outputs most expensive to an organization using AI tooling — subtle logic errors, security vulnerabilities, design misunderstandings that compound downstream — may consume exactly the same credits as correct output. One developer described using 1,180 credits on a single session for results they called "mediocre." The credit cost was visible. Whether those results introduced bugs that would surface in production was not.

Usage-based billing creates a real forcing function for organizational discipline. And when cost is the only metric with infrastructure behind it, cost starts to govern decisions it was never designed to inform. There's already evidence of this in the first week: developers describing how they reduce usage of certain workflows to stay within budget, choosing cheaper models for tasks where they might not be appropriate, defining "good enough" by what's affordable rather than what's correct. Cost became the governing metric by default, because cost is the measurement that's available, and organizations build discipline around what they can see. The meter is precise about consumption. Whether the thing consumed was any good remains, for now, outside the frame.