Run one real task at two effort levels to judge Copilot Fusion

You are deciding whether an effort-level coding agent will improve output without inflating cost. Put one real backlog task through the same test you would give a junior engineer: same prompt, Standard and Max, then diff the results.

Copilot Fusion is generally available in GitHub Copilot as of October 28, 2026, and users select Standard for everyday coding, High for challenging tasks, or Max for the most demanding tasks. It is offered on all Copilot plans in Visual Studio Code, the GitHub Copilot app, GitHub Copilot CLI, and JetBrains IDEs. The benchmark is close on score: Fusion Max cost $5.59 per task and scored 60, while Claude Code with Claude Opus 5.5 Max cost $11.42 and scored 62. The labels are marketing, but the test is mechanical: run the same task twice and let the diff argue.

The same-day A/B keeps the test honest

The checklist is the minimum for a clean comparison. Treat the run like any other task in your queue.

• Pick one real task with a clear done state, such as a failing test, a small refactor, or a bug with a reproducible case.
• Save the prompt, repo state, and acceptance criteria before you run anything.
• Queue the same task at Standard and again at Max, keeping the same prompt and same starting point.
• Save both diffs, both usage reports, and the model names shown in the report.
• Check the usage report for cached context, org policy limits, and any model that appears only in one run.
• Compare the two costs against the cost of the model you already pay for on the same task.

The diff is the real scorecard

Look for the first place the two outputs stop matching. A small change in a helper function may be harmless. A changed migration, a new config flag, or a different error path is a release risk. The benchmark numbers are a prior, not a promise for your repo.

A usage report tells you what happened; a diff tells you whether it mattered. A Max run that changes the public API, adds a dependency, or rewrites a module Standard left alone carries a cost you have to assign. A Standard run that passes the tests while Max only adds comments shows the higher effort level spending tokens on polish.

Usage reports hide the real cost

GitHub charges the models Fusion uses at standard rates, with no separate Fusion fee, and paid Copilot plans get a 30% discount on those model costs through November 30, 2026. Token counts are the visible part of the bill. The hidden part is context. Cached context can make one run look cheaper without a better setting. Org policy may block a model and force a fallback that changes both cost and behavior. A report showing a different model than expected makes the run invalid; rerun it.

Fusion used 7.2 million tokens per task, about half the token usage of Claude Code with Opus 5.5. That spread is what you need to verify on your own workload, because token use depends on how much context your task includes.