GitHub added a per-model token breakdown to the Copilot usage report on 11 August: input, output, cache read and cache write, each mapped to the AI credits it cost.
Four numbers per model is enough to diagnose almost every avoidable cost problem a team has. It is also enough to draw confident wrong conclusions, which is what happens when people read the total and stop.
What each column actually measures
Before any ratio makes sense, be precise about what is being counted.
- Input — everything sent to the model: your prompt, plus whatever context the tool attached. You control the first part and mostly do not control the second
- Output — what the model generated. Usually the most expensive per token, and the one that genuinely correlates with work done
- Cache read — context served from cache instead of being re-sent, typically at a steep discount
- Cache write — the cost of placing context into the cache. You pay this once so that cache read can save you later
The ratios that matter
Absolute numbers tell you the size of your usage. Ratios tell you whether it is healthy.
Cache read against input is the headline. A high proportion of cached input means you are working repeatedly in the same context and getting the discount. A low one means you are paying full price for context you have sent before.
Cache write against cache read is the follow-up. Writes without corresponding reads mean you are filling caches you never reuse — paying the setup cost of an optimisation and never collecting on it.
Output against input is the third. If input dwarfs output by a large multiple, you are spending most of your money on context rather than on generation, which usually means the tool is attaching more than the task needs.
Four patterns and what to do about them
These cover most of what a first read of the report will surface.
- High cache write, low cache read — sessions are too short or too scattered. Do more consecutive work in one repository rather than hopping between many, and the same cache starts paying for itself
- High input, low output, expensive model — an expensive model is answering cheap questions with a lot of context attached. This is the clearest case for routing routine work to a small model such as MAI-Code-1.1-Flash
- One model dominating cost with no matching share of work — check whether it is somebody's deliberate choice or a team-wide default nobody revisited. Defaults are how expensive models become expensive habits
- Output rising much faster than headcount — usually agents, not people. That is not necessarily wrong, but it means your per-seat mental model of cost has stopped applying and needs replacing
What the report cannot tell you
It measures consumption, not value. Nothing in these four columns knows whether the tokens produced a merged pull request or a rejected one, and a cheap month with nothing shipped is not a win.
Pair it with an outcome signal from somewhere else before you conclude anything — merge rate, revert rate, or the effort levels now recorded on Copilot code reviews. Cost data alone reliably pushes teams toward being cheaper rather than being better.
Do not build a per-developer league table
It is the first thing people reach for and it is the wrong instrument. Token consumption tracks the shape of the work: a large legacy refactor burns more than greenfield code, and the person doing it is not being wasteful.
Ranking individuals by token spend teaches people to avoid the expensive tool rather than to use it well, which costs more than any overspend it prevents. Use the report to find structural waste — a bad default, a duplicated automation, an unread cache — and leave individuals out of it.
The fifteen-minute monthly review
Enough to catch nearly everything, and short enough that it actually happens.
- Download the report and sort by credits per model — check the top model is one you chose deliberately
- Check cache read as a share of input, and compare it to last month rather than to any absolute target
- Look for cache write with no matching read, which is the cheapest thing on this list to fix
- Find your most expensive model and ask whether the work it did needed it
- Write down one number to watch next month, and stop there