← Other blogs

Copilot usage by model: why two developers on the same plan diverge

Same prompt, different rate. The model dropdown is a cost decision, and most teams never look at it.

· 2 min read

Copilot usage by model explains most of the gap between two developers on the same plan. One finishes the month comfortably, the other runs out in week three, and neither changed their typing speed. What changed is which model answered.

The same prompt, different cost

GitHub's billing is metered per request, and each model carries its own rate. A request against a large frontier model can consume many times the allowance of one against a smaller, faster model. The editor does not always make that difference obvious, because the dropdown shows names rather than prices.

The practical result is that a habit, such as always picking the most capable model for every inline suggestion, quietly sets the monthly burn. Autocomplete on a frontier model is a different purchase from autocomplete on a small one.

Where to look

Three views help, and none of them is complete on its own.

  • The model picker. It tells you what is available, not what it costs.
  • The account usage page. It shows the allowance and the reset date, but usually not a per-model breakdown.
  • A local tracker. It keeps the model mix over time, which is the view that explains a trend.

On UsageAtlas, usage is recorded per tool and per model on your own machine, so a week of heavy frontier-model work shows up as a shape rather than a single number. It reads usage locally, needs no account, and does not read prompts or responses. Cost figures are estimates, not a bill.

A simple routing rule

Most teams land on something close to this:

Task Model tier Why
Inline completion while typing Smallest available Volume is enormous and the task is shallow
Explaining an unfamiliar file Mid tier Needs context, not genius
A failing test with a long trace Largest One good answer beats five cheap ones
Boilerplate and renames Small or mid Mechanical work with a clear target

The rule is not about saving money for its own sake. It is about spending the allowance where a better answer changes the outcome.

What the numbers will not tell you

A model mix does not say whether the code was good. A cheap month can be a productive one or a quiet one. Nor does it prove that a plan is too small; it may simply be pointed at the wrong tasks.

Read the split alongside the work. If the largest model is handling routine edits, move those down a tier first. If it is handling the hard problems and the month still runs out, the plan is the constraint.

A habit that takes a minute

Once a week, open the tracker and look at two things: the share of requests that went to the top tier, and the day the allowance moved fastest. If both point at the same task type, you have found the setting to change.

That is the whole practice. Models are a cost decision as much as a quality one, and the split is the only place the decision is visible.