Cafiyn Pulse
← Startup Signals
5 min readUpdated 2026-08-31Issue 1

One Copilot code review costs 13 requests. Here is the rest of the maths.

Copilot now bills on usage rather than seats. The headline change is simple. The multipliers underneath it are where teams get surprised.

ByKarthik KumarCafiyn Innovations
At a glance
Changed
All plans moved to usage-based billing on 1 June 2026
Unit
Monthly allotment of GitHub AI Credits
Reasoning models
Can bill at 5x or 20x a standard request
Copilot code review
13x per pull request or IDE review
Auto model selection
10% discount, e.g. 1x becomes 0.9x
Enterprise Cloud
Data residency or FedRAMP adds a further 10%

The change, and the number that matters

All Copilot plans moved to usage-based billing on 1 June 2026, replacing the previous premium-request model with a monthly allotment of GitHub AI Credits.

Per-seat pricing is predictable and usage-based pricing is not, which is the entire story. But the part that catches teams out is not the model itself, it is that a single interaction does not necessarily cost one request.

Copilot code review carries a multiplier of 13. Every time it reviews a pull request, or reviews code in your IDE, your allotment drops by 13. If you wired Copilot review into CI on every PR, work out your monthly PR count and multiply. For a team merging ten pull requests a day, that is a meaningfully different number than the one on the pricing page.

Where the multipliers stack

Advanced reasoning models can consume 5x or 20x a standard request. This is defensible, since a reasoning model genuinely does more work, but it makes model choice a budget decision rather than a preference.

There are discounts too, and they are worth taking. Using auto model selection in Copilot Chat, Copilot CLI, or the Copilot cloud agent earns a 10% multiplier discount, so a model that would bill at 1x bills at 0.9x instead. Letting the platform pick is cheaper than pinning a model, which is the opposite of the instinct most engineers have.

On GitHub Enterprise Cloud, requests processed with data residency or FedRAMP enforcement carry an additional 10% multiplier. If you are in a regulated environment, that is a compliance cost showing up as a line in your AI budget, and it is worth naming it as such when someone asks why the number moved.

Why unattended agents change the risk

Under per-seat pricing, the worst case was bounded: you paid for the seats you had. Under usage-based pricing with agents that run without a human waiting, the worst case is bounded by whatever loop you wrote.

This is the same shape as any metered cloud resource, and the same discipline applies. The reason it surprises people with AI tooling specifically is that the previous model trained everyone to think of it as a fixed cost, so nobody instrumented it.

What to do

  1. 1Count your monthly pull requests and multiply by 13 before deciding whether Copilot code review runs on every PR or only on ones above a size threshold.
  2. 2Turn on auto model selection where you can. The 10% discount applies for doing less work, not more.
  3. 3Treat reasoning models as a deliberate choice per task rather than a default, given the 5x to 20x range.
  4. 4Set a spend alert now. The first genuinely surprising invoice under a metered model is almost always the one nobody was watching for.
  5. 5If you are on Enterprise Cloud with data residency or FedRAMP, budget the extra 10% explicitly so it is not mistaken for drift.

Common questions

When did Copilot billing change?

All plans moved to usage-based billing on 1 June 2026, shifting from premium requests to a monthly allotment of GitHub AI Credits.

How much does a Copilot code review cost?

It carries a model multiplier of 13, so each pull request review or IDE code review reduces your monthly allotment by 13 rather than 1.

Is there a way to reduce the cost?

Auto model selection in Copilot Chat, Copilot CLI, or the cloud agent earns a 10% multiplier discount, so a 1x model bills at 0.9x. Beyond that, the levers are choosing reasoning models deliberately and scoping automated review rather than running it on everything.

Does data residency cost extra?

On GitHub Enterprise Cloud, requests processed with data residency or FedRAMP enforcement carry an additional 10% multiplier.

Sources

Read against the primary documentation rather than secondary coverage. Where a figure comes from a provider's own docs, it is quoted as published on 2026-08-31.

Related

What a code assistant costs per developer →Model your own AI bill →

More from Signals

OpenAI has published seven shutdown dates. Here is the calendar.
Seven dates between 23 October and 26 February. Every one announced months ahead, in public. The failure mode is never the notice.
Nvidia bought Hugging Face. Your open-weight fallback is now a vendor relationship.
The commitments say nothing breaks. The structural change is that "we can always self-host" is now a dependency on one vendor rather than on a neutral commons.
The Videos API goes away on 24 September and OpenAI names nothing to move to.
Every other deprecation on OpenAI's page offers a migration target. This one does not, which makes it a different kind of problem.
Whisper shuts down on 26 February. The risk is not the deadline.
Five months is a generous window. Whisper is also typically wired into a background job that fails into a queue rather than into a user's face.
The Assistants API is gone. Your app may not have told you yet.
It shut down on 26 August 2026. The replacement is not a drop-in, and the most common migration mistake changes your model's behaviour without throwing an error.
Row-level security is on by default now. That is not the same as being protected.
Lovable fixed the default. The gap that actually leaks data was never the default, it was the policy, and no platform can write that for you.
Your model comparison is measuring the wrong number.
Every price list is per token. Your bill is per finished piece of work. Those two numbers can rank models in opposite orders.
An agent with production log access is a new identity. Treat it like one.
Always-on agents that read production logs are arriving fast. The access-control question they raise is old and well understood, which is the good news.

Open the tool.

Ten production categories, three minutes, and a ranked list of what to fix first.

Check your own stack