The short version
GitHub now picks the model for you, and says it does so at about half the cost of running the top model alone. Copilot Fusion became generally available on October 28, 2026, alongside a new Copilot app, multi-agent Projects and AI-credit billing for GitHub Enterprise. For companies that buy custom software development, the practical shift is from paying per seat to paying per task, which makes cost per finished change the number to watch.
What did GitHub announce at Universe 2026?
The centerpiece is Copilot Fusion, a service that handles both model selection and orchestration. According to GitHub, it looks at what a task needs across reasoning, code generation, debugging and tool use, then decides how to run it across one or more models. It builds on Project HydraFusion, the multi-model routing research GitHub previewed in September, combined with the existing Copilot Auto Model Selection. VentureBeat reports that Fusion is available in VS Code, the Copilot app, Copilot CLI and JetBrains IDEs.
Around it, GitHub shipped a new Copilot app built for agentic work across projects, Copilot Projects for running parallel agent sessions across several repositories (public preview), session continuity between the app, VS Code, CLI and web, Copilot Automations, and a new Code Review Max tier. In VS Code, the Agent Host that keeps sessions running outside an editor window is now generally available, and GitHub published an open Agent Host Protocol with client libraries for TypeScript, Rust, Go, Kotlin and Swift. A computer use feature that lets agents operate desktop apps is in public preview in the Copilot app and experimental in VS Code, with per-app approval and an organization-level off switch.
How good is Copilot Fusion, really?
GitHub’s headline claim is that Fusion Max scores within two points of Claude Opus 5.5 Max at half the cost and 40% less time per task. VentureBeat, citing the Artificial Analysis Coding Agent Index, put numbers on it: Fusion Max scored 60 against roughly 62 for Claude Code on Opus 5.5 Max, at an average of $5.59 per task versus $11.42, about 51% less, using around half the tokens. The model pool includes Claude Opus 5.5, Claude Sonnet 5.5, GPT-6 Astra and GPT-6 Sol.
The results are uneven. Fusion Max did well on Terminal-Bench 4.0 (71) but scored 51 on FrontierCode, and benchmark runs leave out human review, failed attempts and security work. Microsoft CEO Satya Nadella quoted “up to 77% lower cost,” which is Microsoft’s own figure and separate from the independent 51% result. Treat the savings as a ceiling to test, not a forecast.
What changes in how Copilot is billed?
GitHub says Enterprise customers can now use every Copilot feature without buying a separate Copilot subscription and pay only for the AI credits they consume. It is available through GitHub sales now, with self-serve promised later. VentureBeat also reports a temporary 30% discount on model costs for paid subscribers. Combined with Fusion, this pushes Copilot spending from a fixed per-seat line toward a variable cost that depends on how many tasks agents run and at what effort level.
What it means for US & EU software teams
Agents are already doing a large share of the work. GitHub’s Octoverse data, presented at Universe, shows that 40% of pull requests began as agent drafts by September, up from 17% in March. When cheaper routing makes each task less expensive, teams will run more of them, so review capacity and test coverage become the real limits. Budget for reviewers and automated checks, not just for credits.
Routing also raises a governance question. If a router can send parts of a task to models from different vendors, your security and legal teams need to know which models and regions can see your code, and whether your contracts and data processing terms cover all of them. For EU clients and regulated work in FinTech or HealthTech, ask for that list in writing, keep logs of which changes were agent-generated and who approved them, and switch off computer use until you have a policy for it.
What should teams do now?
- Pilot on your own code. Run Fusion and your current agent on the same backlog for two weeks and compare cost per merged pull request, rework and review time.
- Set credit budgets per team. Usage-based billing needs limits and alerts; decide which tasks may use the Max effort level.
- Map the model pool. Confirm which models and hosting regions Fusion may use for your organization and check them against your data rules.
- Keep humans on risky paths. Authentication, payments, data migrations and infrastructure changes should always get a named reviewer.
- Lock down computer use and plugins. Use organization policies to disable desktop control and unapproved marketplace plugins until reviewed.
- Avoid hard lock-in. Keep prompts, agent instructions and test suites in your repository so you can change tools if prices or quality shift.
Frequently asked questions
What is GitHub Copilot Fusion?
Copilot Fusion is a model selection and orchestration service that GitHub made generally available at Universe 2026 on October 28. It combines runtime orchestration from Project HydraFusion with Copilot Auto Model Selection, so Copilot decides which models handle which parts of a coding task. Developers choose an effort level: Standard, High or Max.
How does Copilot Fusion compare with Claude Code on cost?
On the Artificial Analysis Coding Agent Index, Fusion Max scored within two points of Claude Code running Claude Opus 5.5 Max, at an average of $5.59 per task versus $11.42, about 51% less, and in about 40% less time, VentureBeat reported. Results were uneven across benchmarks, and the figures do not include human review, retries or security work.
What changed for GitHub Enterprise customers?
GitHub says Enterprise customers can now use all Copilot features without a separate Copilot subscription and pay for the AI credits they consume. The option is available through GitHub sales now, with self-serve sign-up announced as coming soon.
Should teams switch their coding agents to Copilot Fusion?
Not on benchmark numbers alone. Run a short pilot on your own repositories, compare cost per merged pull request and rework rate against your current tool, check which models and regions touch your code, and set credit budgets and approval rules before rolling it out widely.
Sources
GitHub — GitHub Universe 2026: Evolving every layer of the stack (primary source, October 28, 2026)
GitHub Changelog — GitHub Copilot in VS Code, Universe releases (October 28, 2026)
VentureBeat — GitHub Copilot Fusion mixes models to achieve near-Opus 5.5 coding performance at half the per-task cost (October 28, 2026)
The Next Web — GitHub Universe: Copilot Fusion launches as agents start 40% of PRs (October 28, 2026)
The New Stack — GitHub debuts Copilot Fusion, a new model selection and orchestration system (October 28, 2026)