Copilot Fusion routes coding tasks, but benchmark savings need a local test

GitHub made Copilot Fusion generally available on October 28, 2026. It lets Copilot choose models and coordinate them for a coding task while the developer selects an effort level.

GitHub describes Fusion as a system that routes and orchestrates models for each prompt, rather than as a new model. That can reduce model-picking work, but it does not turn a benchmark result into a forecast for a team's bill or code quality.

What the benchmark says, and what it does not

Artificial Analysis' Coding Agent Index v2.0 gives Copilot Fusion Max a score of 60 and Claude Code with Claude Opus 5.5 Max a score of 62. The same page lists average API cost per task at US$5.59 for Fusion and US$11.42 for Claude Code.

The test also separates the tasks. Fusion scores 51 on FrontierCode v1.1, 71 on Terminal-Bench 4.0 and 57 on SWE-Atlas-QnA-AA. Claude Code scores 56, 63 and 66. The combined score hides those differences, so a team should map its work to command-line operations, repository changes or question answering before drawing a conclusion.

These are complete agent configurations with models, tools and execution behavior. They are not measurements of one model in isolation. The API figures use the benchmark's task mix and pricing assumptions; they do not include a team's code review, failed runs, waiting time, deployment work or manual repair.

Measure	Copilot Fusion Max	Claude Code with Opus 5.5 Max
FrontierCode v1.1	51	56
Terminal-Bench 4.0	71	63
SWE-Atlas-QnA-AA	57	66
Composite score	60	62
Average API cost per task	US$5.59	US$11.42

How Fusion routes and hands off work

GitHub's documentation defines Fusion as a runtime router and orchestrator that selects models and execution patterns for each task. The developer chooses Standard, High or Max effort; Fusion can then decide whether one model is enough or whether another model should review, reject or revise the first response.

Single sends the task to one model. Cascade lets an efficient model draft a solution and a judge decide whether to accept it or escalate. Critique has one model draft while other models review in parallel, after which the original model revises its answer. Two prompts in the same session can therefore use different patterns.

That handoff is only as good as the context passed between steps. A request that omits the files in scope, compatibility constraints, tests and completion conditions can produce a plausible first change and a plausible review of the wrong assumption. More model calls add another chance to inspect the work, not a guarantee that the missing requirement will be found.

GitHub's documentation also warns that an intermediate draft may already have edited the workspace when Fusion discards it. Before accepting a response, compare the version-control diff, rerun the relevant tests and check that every remaining change belongs to the task.

Why the API comparison is not a team budget

GitHub bills Fusion for the models it uses at their standard rates and does not add a separate Fusion fee. High and Max can call more than one model, so the same task may consume more AI credits than Standard.

GitHub says one AI credit equals US$0.01 and that usage depends on the model and tokens consumed. A long coding-agent session can use more credits than a short chat request. Included credits, pooled credits, spending limits and charges after the allowance are controlled by the account's plan.

For individual plans, GitHub's current public table lists Copilot Pro at US$10 per month, Pro+ at US$39 and Max at US$100, with different monthly credit allowances. The plan documentation also says that Business and Enterprise credits are pooled and that use beyond the pool is billed separately. These are published USD plan figures, not a promise about taxes, payment conversion or an organization's negotiated contract.

The GitHub Changelog says paid plans receive a 30 percent Fusion model-cost discount through November 30, 2026. That date matters: a pilot that starts before the deadline should record the discount, plan, model policy and usage date rather than treating the observed price as permanent.

How this plays out outside Taiwan

A developer in the United States, Singapore, the United Kingdom or Australia can start with the same product question: is Fusion available on the account, in the surface and at the effort level the team intends to test? GitHub says Fusion is available with all Copilot plans and lists Visual Studio Code, Copilot CLI, the GitHub Copilot app and JetBrains IDEs as generally available surfaces. That does not mean every plan exposes every effort level or every model. GitHub says model choice is subject to the subscription type and administrator policies.

For an individual in the United States, the public plan table is expressed in US dollars: Pro is listed at US$10 per month, Pro+ at US$39 and Max at US$100. The practical check is the account's GitHub billing page, because the public table does not by itself show sales tax, a card issuer's currency conversion, a promotional price or whether an employer already pays for access. A developer should record the displayed plan, the included AI-credit allowance and the budget for extra usage before comparing Fusion with a separately billed coding agent.

For a Singapore developer, the same USD table is useful for comparing list prices, but it is not a Singapore-dollar invoice. GitHub's billing documentation defines AI credits in USD and describes the usage unit as one credit for US$0.01 of metered use. The pilot should therefore save both the GitHub usage report and the payment record, then compare the converted charge with the benchmark's US$5.59 figure. A lower USD API estimate cannot answer whether the final card charge is lower after conversion, taxes or a company purchasing arrangement.

For a United Kingdom team, plan access and model policy should be checked by the organization owner rather than inferred from a developer's personal account. GitHub's plan rules distinguish individual plans from Business and Enterprise plans and say that organizations can centrally manage access, policies and usage controls. The local procurement record should show whether the team has seats, pooled credits or metered access, who may change the model policy, and where an administrator sets the budget. The useful UK-specific question is not whether a US dollar is a good price; it is which account is responsible for the next AI-credit charge.

For an Australian team, the availability check is similar but the operating boundary may be different. GitHub says Fusion can choose only models allowed by the plan and administrator policies, including policies that restrict access to data-resident or FedRAMP-compliant models. If an Australian organization has a regional data or supplier rule, the team should ask its GitHub administrator which model pool is enabled and whether that policy removes Fusion from the picker. The result can be documented from the account configuration rather than guessed from the existence of the public feature page.

Across all four markets, GitHub's billing unit changes how the pilot should be run. Individual credits reset monthly and unused credits do not carry over, while organization and enterprise plans can pool allowances and charge separately after the pool is exhausted. A team should set a spending limit before it gives Fusion the hardest tasks, export usage after each test batch, and keep Standard, High and Max results separate. Otherwise a short trial can mix a temporary discount, a pooled enterprise allowance and different model calls into one number that cannot be reproduced.

Availability also changes when a company policy changes. GitHub says Fusion will not use models excluded by the subscription, administrator settings, data-residency or FedRAMP restrictions, or evaluation-model policies. If no eligible model remains, Fusion does not appear in the model picker. The pilot record should therefore include the plan country or billing entity, the organization policy, the enabled IDE or CLI surface and the model list shown on the test date.

Finally, local purchasing and product access are only part of the result. The software team should run the same representative tasks under the same repository revision, prompt, tests and reviewer. Report API credits or metered spend in USD, report the actual local charge separately, and count review minutes, failed tests, reverted edits and reruns. This gives a US, Singapore, UK or Australian team a defensible local result without pretending that one country's price or policy applies everywhere.

A small pilot can expose the real cost

Use one task set that reflects normal work rather than a demonstration chosen to flatter the tool. Include a small edit, a difficult bug, a cross-file change and a task with a clear test failure. Keep the repository revision, prompt, model policy and completion conditions fixed.

1. Separate task types. Label work as routine, debugging, command-line or cross-file so the composite average does not hide a weak category.
2. Run the effort levels separately. Start with Standard, then test High or Max only where the task needs deeper checking. Record every additional model call or revision round that the interface exposes.
3. Track the whole bill. Record AI credits, USD metered cost, local payment amount, waiting time, review minutes, failed tests, reverts and manual edits.
4. Inspect the handoff. Check the files changed, assumptions made, tests run, failed checks and unresolved questions. Assign one person to accept or reject the final diff.
5. Include failures. Keep incomplete and heavily repaired tasks in the average. Removing them is a direct way to understate cost.

GitHub and Accenture reported higher pull-request merge rates and build success in a randomized enterprise study of Copilot users, but that evidence concerns the study's participants and setup. GitHub's research page reports a 15 percent increase in pull-request merge rate and an 84 percent increase in successful builds. It does not establish what Fusion will do for another team.

Microsoft's KOHLER customer story reports that 5,000 employees saved two to five hours a week, while code releases rose 25 percent and error defects fell 75 percent. Microsoft presents those figures as a customer case study tied to KOHLER's training and change-management program, not as a controlled estimate for Fusion. It is a reminder to measure adoption and operating practice alongside model cost.

Questions to settle before adoption

Is Copilot Fusion a new AI model?

No. It is a routing and orchestration system that selects models and execution patterns at runtime.

Does every Fusion task use multiple models?

No. Single can use one model. Cascade and Critique add judging, review or revision when the selected workflow calls for it.

Is Fusion always cheaper?

No. The benchmark lists a lower average API cost in its test setup. Real spending depends on the task, model calls, effort level, plan, credits, discounts and later repair work.

What should an enterprise check first?

Confirm the enabled plans and models, the administrator policy, the AI-credit budget, the actual workspace diff and the tests that define acceptance. Then compare Fusion with the existing workflow on the same representative tasks.