# GitHub's Copilot Fusion routes coding work across models

Copilot Fusion now routes coding work across one or more models, while Copilot Projects can coordinate parallel sessions across repositories.[1][2] Independent tests found near-Opus 5.5 composite performance at about half the measured task cost, but results varied by benchmark.[6]

GitHub used Universe 2026 to move Copilot beyond a single assistant session.[1] Fusion is generally available and can choose Single, Cascade or Critique execution at three effort levels.[2] GitHub also put managed parallel sessions across repositories into public preview through Projects, and eligible GitHub Enterprise Cloud customers working through sales can use Copilot features without a separate per-seat Copilot subscription, with AI usage metered from the first request.[4] Teams should test accepted changes, review time, policy enforcement and total AI-credit spend before changing defaults.

## What GitHub released at Universe

[GitHub's Universe announcement [1]](https://github.blog/news-insights/company-news/github-universe-2026-evolving-every-layer-of-the-stack/) presents Copilot as a connected software-delivery system. GitHub announced Fusion, cross-repository Projects, session continuity, cloud automations and a Max code-review tier.

[The Fusion release notes [2]](https://github.blog/changelog/2026-10-28-route-and-orchestrate-models-with-github-copilot-fusion/) say every Copilot plan can use Fusion in Visual Studio Code, the Copilot app, Copilot CLI and JetBrains IDEs. Users choose Standard, High or Max effort instead of choosing one model.

Fusion can send a task to one model, start with an efficient model and escalate after a quality check, or ask one model to critique another model's draft before revision.[2] Administrators can restrict the eligible model pool through enterprise and organization policies.[2]

## What changed beyond model routing

[GitHub's Copilot app release [3]](https://github.blog/changelog/2026-10-28-github-copilot-app-universe-releases/) puts Projects in public preview. A project can combine repositories, design material, Slack threads and notes, then run several agents and return outputs such as pull requests for review.

The app also adds public-preview computer use on macOS and Windows. GitHub says Copilot asks before controlling an app, and administrators can disable the feature.[3] Teams still need to test what the approval covers and whether connected data stays within their intended boundary.

[GitHub's enterprise access notice [4]](https://github.blog/changelog/2026-10-28-github-enterprise-now-includes-access-to-github-copilot/) introduces another purchasing route for GitHub Enterprise Cloud customers working with a GitHub or Microsoft account team. Copilot access can be included for licensed users, but usage is metered against the enterprise budget from the first request and no monthly AI credits come with each seat.

## What the independent benchmark supports

[Artificial Analysis tested Fusion Max [6]](https://artificialanalysis.ai/articles/github-copilot-fusion-orchestrates-anthropic-and-openai-models-to-balance-performance-and-cost) in Copilot CLI on its Coding Agent Index v2.0. Fusion scored 60 and averaged $5.59 per task, while Claude Code with Opus 5.5 Max scored about 62 and averaged $11.42 per task.

The three component scores varied. Fusion scored 71 on Terminal-Bench 4.0, 51 on FrontierCode and 57 on SWE-Atlas-QnA-AA.[6] A team whose work resembles one component may see a different trade-off from the headline average.

Artificial Analysis measured roughly 7.2 million tokens per task for Fusion, about half the Opus 5.5 Claude Code configuration in the same report.[6] GitHub says Fusion took about 40% less time than the compared Claude Code configuration.[2] Artificial Analysis does not publish that timing comparison.[6]

## A rollout test for coding teams

- Choose 20 to 40 representative tasks across bug fixes, repository questions, refactors, tests and tool use. Keep expected outcomes and acceptance checks fixed.
- Run the same tasks with the current model choice and Fusion Standard, High and Max. Use clean branches or isolated worktrees so one run cannot influence another.[2]
- Record whether each change passes tests, survives human review and reaches the intended outcome. Count retries, abandoned runs and reviewer minutes.
- Export AI usage and calculate credits per accepted change rather than relying on cost per prompt. Include Actions or cloud-sandbox consumption where the workflow uses them.
- Confirm that disabled models never appear in Fusion's eligible pool. Remove the available budget in a test account and verify that GitHub blocks Chat, code review and cloud Copilot as documented.[2][4]
- Test Projects on work that crosses repositories. Require separate checks, ownership and approval for every repository before merge.[3]
- Keep computer use off unless the task requires a desktop app. When enabled, test approval prompts, screen access, connected accounts and revocation.[3]
- Set a rollback rule before changing the default. Return to explicit model selection if acceptance falls, review time grows or spend exceeds the agreed limit.

## Security and governance checks

[GitHub also opened AI-application scanning in public preview [5]](https://github.blog/changelog/2026-10-28-code-scanning-for-ai-applications-in-public-preview/) for repositories with GitHub Advanced Security and the feature enabled. GitHub says it can flag code paths involving indirect prompt injection, sensitive information disclosure and excessive agency on pull requests.

Treat the scanner as one review input. GitHub has not published recall, precision or a complete coverage map for these AI-specific detections in the cited release.[5] Continue threat modeling tool permissions, untrusted inputs, secrets, network access and human approval paths.

Teams also need to revise supplier review for dynamic routing. Administrators should document the models allowed in Fusion, the data rules attached to each provider, the default effort level and the evidence available after a request.[2]

## What remains uncertain

- GitHub does not promise that Fusion will cost less or finish faster on every workload.[2]
- The public materials do not disclose every routing rule, quality threshold or per-request model choice made by Fusion.[2]
- The independent test covers Fusion Max in Copilot CLI and three benchmark components. It does not establish results for every IDE, repository, language or effort level.[6]
- Projects and computer use remain public previews, so behavior, availability and interfaces can change.[3]
- The new enterprise access route is not self-serve yet, includes no per-seat AI-credit allowance and does not replace existing Copilot Business or Enterprise plans.[4]
- GitHub reports that code review Max led ReviewBench on high-severity F1, but the Universe post does not provide a full score table or enough detail for an independent comparison.[1]

## Frequently asked questions

- Is Copilot Fusion a foundation model? No. GitHub describes it as a system that selects and coordinates eligible models for each task.[2]
- Can an administrator block a model? Yes. GitHub says enterprise administrators can set model policy for Fusion separately and choose a default Fusion effort level.[2]
- How is Fusion billed? GitHub bills each underlying model call at that model's standard rate and currently offers paid plans a temporary 30% discount on Fusion model costs.[2]
- Does the benchmark prove lower production cost? No. It measures a defined coding index. Teams still need to count failed attempts, review, integration, security work and accepted outcomes.[6][7]
- Can Projects work across repositories? GitHub says Projects can run parallel managed sessions with context and outputs spanning several repositories.[1][3]
- Does GitHub Enterprise now include free AI usage? No. Eligible customers can receive feature access without a separate Copilot subscription, but usage is metered from the first request and requires an available budget for Chat, code review and cloud Copilot.[4]

## Sources

- [[1] GitHub: Universe 2026 announcement](https://github.blog/news-insights/company-news/github-universe-2026-evolving-every-layer-of-the-stack/)
- [[2] GitHub: Copilot Fusion release notes](https://github.blog/changelog/2026-10-28-route-and-orchestrate-models-with-github-copilot-fusion/)
- [[3] GitHub: Copilot app Universe releases](https://github.blog/changelog/2026-10-28-github-copilot-app-universe-releases/)
- [[4] GitHub: Enterprise Copilot access and billing](https://github.blog/changelog/2026-10-28-github-enterprise-now-includes-access-to-github-copilot/)
- [[5] GitHub: AI application code scanning preview](https://github.blog/changelog/2026-10-28-code-scanning-for-ai-applications-in-public-preview/)
- [[6] Artificial Analysis: Copilot Fusion benchmark](https://artificialanalysis.ai/articles/github-copilot-fusion-orchestrates-anthropic-and-openai-models-to-balance-performance-and-cost)
- [[7] VentureBeat: Copilot Fusion analysis and GitHub interview](https://venturebeat.com/orchestration/github-copilot-fusion-mixes-models-to-achieve-near-opus-5-5-coding-performance-at-half-the-per-task-cost)