A cheap model drafts; a stronger one reviews. That sounds economical until most drafts need another full attempt. Evaluating Copilot Fusion requires the cost of that entire path, including rejected work.
In its October 28 GA announcement, GitHub reports a Coding Agent Index score of 60 for Fusion Max, within two points of Claude Code with Opus 5.5 Max, at about half the average API cost and tokens and roughly 40% less time per task. These are GitHub's reported comparison results, not CodeBridge measurements or universal savings. This review is dated October 29.
Treat the selection as an execution policy
The Fusion documentation describes three patterns:
| Pattern | Work performed | Additional variable to measure |
|---|---|---|
| Single | One model completes the task | Overhead on straightforward work |
| Cascade | Draft, judge, escalate if needed | Escalation frequency and judging cost |
| Critique | Draft, review by other models, revise | Errors found and regressions introduced |
The runtime chooses the composition within your effort level and available models. Organization policies also affect that pool. Record effort, client version, and observed models and patterns alongside the name “Fusion.”
Review by a different model is not an independent correctness oracle. Two models can share the same missing requirement. External tests and human examination of the requirements remain the useful yardsticks for whether another review pass earned its cost.
A small calculation with a large consequence
Every price below is invented for instruction, not a model rate. This code was executed locally with Node.js; it needs Node 18+, no packages, keys, or network. It makes no model calls.
const draft = 0.20, judge = 0.05, repair = 1.00;
for (const escalation of [0.10, 0.80]) {
const expected = draft + judge + escalation * repair;
console.log(`escalation=${escalation.toFixed(2)} expected_usd=${expected.toFixed(2)}`);
}
const attempts = [{ cost: 0.35, ok: true }, { cost: 0.35, ok: false }];
console.log(`cost_per_success_usd=${(attempts.reduce((n, r) => n + r.cost, 0) / attempts.filter(r => r.ok).length).toFixed(2)}`);
escalation=0.10 expected_usd=0.35
escalation=0.80 expected_usd=1.05
cost_per_success_usd=0.70
Assume a direct stronger-model attempt costs $1. A 10% escalation rate wins in this simplified model; an 80% rate loses. Two $0.35 attempts with only one success cost $0.70 per successful completion. Dropping the failed attempt from the accounting conceals real expenditure.
Under these assumptions, the break-even escalation rate is 75%: solving 0.20 + 0.05 + p × 1.00 < 1.00 gives p < 0.75. Meeting that cost condition still needs a quality constraint, because lower success rates can increase cost per successful completion.
Caching, multiple critics, retries, tools, and human repair time are absent from this calculation. Use it to decide what to log, not to predict an invoice. With zero successes, report the metric as undefined rather than $0.
Compare from identical starting states
Separate small bug fixes, multi-file refactors, and explanation-only questions. Run the fixed-model baseline and Fusion in independent working directories from the same commit. Otherwise an earlier attempt's edits or unrecorded cache conditions can change the task given to the next configuration.
| Record | Why it matters |
|---|---|
| Task ID, starting commit, prompt, effort, policies | Establish comparable inputs |
| Requirements met, tests, reviewed diff | Distinguish a completed response from a correct result |
| Tokens, cost, failures, retries for all attempts | Calculate total cost per success |
| Completion time and human repair time | Test whether API speed translates into saved work |
Alternate execution order and repeat runs to reduce time-of-day and transient service effects. With a small sample, show individual task outcomes as well as averages. A favorable mean can hide a regression on the tasks that matter most.
The docs warn that discarding a draft does not automatically undo edits already made to the workspace. Keep runs isolated and inspect their diffs. No Copilot run or task comparison was performed for this article.
Separate the promotion from the efficiency
Billing sums the models used at their standard rates. The documentation checked on October 29 advertises a 30% Fusion model-cost discount for paid plans through November 30, 2026. Log both discounted and undiscounted costs so the comparison survives the promotion. Do not multiply the benchmark's “half” by the promotion and present it as an established saving.
Our earlier HydraFusion article describes a preview snapshot. For GA, judge adoption by verified completions and total failed-work cost on your workload. Interactive questions put greater weight on latency; delegated coding work puts greater weight on correct, reviewable output.
The paid Java/Spring Copilot course below offers practice in the underlying coding and testing workflow. It is not advertised here as a Fusion-specific curriculum. The link is promotional.