← FinOpsAid blog

What actually consumes AI credits? The Copilot surfaces that cost you money — FinOpsAid blog

What actually consumes AI credits? The Copilot surfaces that cost you money

Completions are free, agents are not. A practical map of which Copilot surfaces burn AI credits, why model choice multiplies the bill, and where the cheap levers are.

Now that Copilot consumption is metered, the question every engineering leader eventually asks is disarmingly practical: what is actually costing us? It's the right question, and the answer is narrower than most people assume — which is good news, because it means the levers are few and specific.

The short version: the things developers do most often are free, and the things they do occasionally can be expensive. That inversion catches people out, because it means your credit burn correlates poorly with how much your team likes Copilot.

The free surfaces

Code completions and Next Edit suggestions consume no credits. They're included in every plan, unmetered, however much your developers use them.

This is the single most important fact about managing Copilot cost, and it's routinely missed. The classic Copilot experience — ghost text appearing as you type, accepted with a tab — is the thing that drove adoption, generates most of the perceived value, and contributes nothing to your credit consumption. A developer who lives in completions all day costs you exactly their subscription and not a cent more.

It also means the old adoption metrics have quietly changed meaning. Acceptance rate and suggestion volume still tell you whether the tool is landing. They no longer tell you anything about what it costs.

The metered surfaces

Credits are consumed by the surfaces where a model does substantial work on your behalf: Copilot chat, the coding agent, and code review.

The economics differ sharply between them. A chat exchange is bounded — a question, some context, an answer. An agent task is not: it reads files, plans, edits, runs things, re-reads, and each of those steps is tokens. A single agent run on a large repository can consume more than a developer's entire month of chat.

That's not an argument against agents. It's an argument for knowing which of your teams have adopted them, because the difference between an org that uses chat and an org that uses agents is not a percentage — it's a multiple.

Tokens are the unit, and cached tokens count

Credits are consumed against token usage: input, output, and cached tokens, each at the published rate for the model that served the request.

Two consequences worth internalising. First, input matters as much as output. A prompt that drags in a large file, a long conversation history, or a wide swathe of repository context costs before the model has written anything. Verbose context is a cost decision, not just a quality one.

Second, the model is a price multiplier. The same task routed to a frontier model versus a smaller one differs in cost by a large factor, because the rates differ by a large factor. GitHub added several models to Copilot in July 2026 alone — Claude Opus 5, Gemini 3.6 Flash, three GPT-5.6 variants, Kimi K2.7 — and each carries its own rate. Model selection has become one of the most direct cost levers you have, and in most organizations it's set by individual preference with no guidance at all.

Which Copilot surfaces consume AI credits Code completions and Next Edit suggestions are included in every plan and consume no AI credits regardless of volume. Copilot chat, code review and the coding agent consume credits against token usage, with the coding agent consuming the most per task because it reads, plans, edits and re-reads across many steps. FREE — UNMETERED, ANY VOLUME Code completions · Next Edit suggestions the highest-volume surface, and it costs nothing beyond the subscription METERED — CREDITS AGAINST TOKEN USAGE Chat bounded exchange Code review scales with diff size Coding agent read, plan, edit, re-read — unbounded AND THEN MULTIPLIED BY Model rate — a frontier model and a small one differ by a large factor for the same task Context size — input and cached tokens are charged, so a wide prompt costs before it answers Your credit burn tracks which surfaces you use, not how much your team likes Copilot.
The volume surface is free and the occasional surface is metered. That inversion is why adoption rate no longer predicts cost, and why two teams with identical acceptance rates can produce very different bills.

Where the cheap levers are

Set a default model, per use case. Not a mandate — a default. Frontier models for genuinely hard problems, cheaper ones for routine chat and small edits. Most developers have no idea the choice has a price attached, and will happily take a sensible default.

Watch agent adoption as a cost event. When a team starts using the coding agent, their consumption curve changes shape. That's worth knowing at the time rather than at the invoice.

Treat context width as a cost. Prompts and configurations that pull in huge amounts of repository context are paying tokens for material the model may not need.

Don't optimise the free surface. Any effort spent discouraging completions is effort spent reducing value at zero cost saving. This sounds obvious and is nonetheless a mistake organizations make within weeks of seeing their first credit bill.

Watch consumption against allowance, per user. The ratio, not the total — it's the only figure that tells you whether anything needs doing.

The organizations that manage this well won't be the ones that used Copilot least. They'll be the ones that knew which surface their spend came from, and set a default before the invoice set one for them.

Both halves are visible in FinOpsAid: Copilot tracks credit consumption against allowance per user and per team alongside adoption, with model mix broken out by tier — so a jump in spend resolves to a surface and a model rather than a total, reconciled against the billed invoice rather than reported in its place.

Frequently asked questions

Do GitHub Copilot code completions cost AI credits?

No. Code completions and Next Edit suggestions are included on every plan and consume no credits, however heavily they are used. The classic ghost-text experience that drove Copilot adoption contributes nothing to credit consumption — a developer who lives in completions costs exactly their subscription.

Why doesn't our Copilot adoption rate predict our Copilot cost?

Because the volume surface is free and the occasional surface is metered. Acceptance rate and suggestion volume still tell you whether the tool is landing, but they no longer tell you what it costs. Two teams with identical acceptance rates can produce very different bills.

Do cached and input tokens count toward AI credits?

Yes. Credits are consumed against input, output and cached tokens alike, each at the published rate for the model that served the request. That makes prompt context a cost decision: pulling in a large file, a long conversation history or wide repository context costs before the model has written anything.

Why is a Copilot agent run so much more expensive than a chat message?

A chat exchange is bounded — a question, some context, an answer. An agent task is not: it reads files, plans, edits, runs things and re-reads, and every step is tokens. A single agent run on a large repository can consume more than a developer's entire month of chat.

Want to see which surfaces and models your credits are going to? Connect your GitHub org — read-only and free during beta — or explore the demo dashboard first.