Copilot's usage report now shows your cache writes

GitHub shipped a small billing change today that fixes a real problem: the AI usage report now breaks tokens out per model. For each model you used, you see input, output, cache read, and cache write tokens next to the AI credits they consumed. If you've been running GitHub Copilot on usage-based billing and couldn't explain a line item, this is the missing column.

What actually happened

Per the GitHub Changelog, the per-model token breakdown is live in the AI usage report, downloadable from the AI usage page in billing settings. It's available to admins on Copilot Business and Copilot Enterprise, and to all users on Copilot for individuals. GitHub's stated purpose is to let you trace how each token type contributes to a model's cost.

Cache write is the interesting one. Prompt caching is normally sold as a discount — cached reads are cheap. What gets skipped in that pitch is that writing the cache costs more than an ordinary input token. On a workload with high prompt churn, where the agent rebuilds context every turn instead of reusing it, you pay the write premium repeatedly and never collect the read discount. Until today that failure mode was invisible inside an aggregate credit total. Now it's a column you can sort.

This lands on top of the governance layer GitHub built after Copilot moved to usage-based billing on June 1 at one AI credit per cent — per-user budgets on cost centers, OpenTelemetry export, and BYOK.

Why it matters for your business

You cannot reduce a cost you cannot decompose. A team that only sees "4,200 credits last month" has exactly one lever: use Copilot less. A team that sees cache-write tokens dominating on one model and one repo has a specific, cheap fix — usually a prompt or agent-config change, not a headcount conversation.

Pull the report this week and look for three patterns. One: a model whose cache-write volume rivals its input volume, which means your context isn't being reused and something upstream is thrashing it. Two: output tokens far exceeding input on routine tasks, which usually means an agent is generating whole files where a diff would do. Three: an expensive model handling work a cheaper one would finish — the per-model breakdown makes model routing an evidence-based decision rather than a preference.

The broader habit is what matters. Every metered AI tool you adopt will eventually hand you a bill you can't explain unless you demand attribution up front. Ask for the per-model, per-token-type breakdown before you scale seats. Vendors who can't produce it are asking you to fly blind on a variable cost.

Sources: GitHub Changelog — per-model token breakdown in the usage report, The GitHub Blog — Copilot moving to usage-based billing.