GitHub made Copilot Fusion generally available on October 28, 2026, at GitHub Universe in San Francisco, routing models by task so developers skip manual selection. GitHub says Fusion Max lands near a leading coding agent at about half the average API cost per task, alongside a new Copilot app, cross-repo projects, and an infrastructure expansion.
Key Takeaways
- Copilot Fusion became generally available on October 28, 2026.
- Fusion Max scored 60, within two points of Claude Code with Opus 5.5 Max.
- GitHub says that run used about half the API cost and about 40% less time per task.
- Paid plans get a limited-time 30% discount on Fusion model costs.
What Copilot Fusion Changes in Daily Coding
Selecting Copilot Fusion chooses a routing system, not a single model, GitHub’s Universe post explains. Each prompt is judged on reasoning, code generation, debugging, and tool use, then run as one model or several. The system joins Project HydraFusion, previewed on September 4, 2026, with Copilot’s auto model selection.
GitHub describes three execution workflows:
- Single: one model solves the task.
- Cascade: an efficient model drafts, then a quality check accepts it or escalates.
- Critique: one model drafts, a second reviews, and the first revises.
Developers still pick effort: Standard for everyday coding, High for deeper reasoning, Max for the hardest work. Admins can limit Fusion’s model list, disable a model for Fusion only, and set any effort level as the default. Users are told to “choose Copilot Fusion and let Copilot intelligently route and orchestrate which models it uses for you.”
How Copilot Fusion Max Compares on Cost
GitHub says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code with Opus 5.5 Max, The New Stack reported. That result used about half the average API cost per task, about half as many tokens, and about 40% less time per task. The index figures are GitHub’s.
| Measure | Fusion Max, per GitHub |
|---|---|
| Coding Agent Index score | 60, within two points of Claude Code with Opus 5.5 Max |
| Average API cost per task | About half |
| Tokens per task | About half as many |
| Time per task | About 40% less |
Unite.AI reported that September preview policies improved verified quality by 4.9 percentage points at 67% lower estimated cost than Claude Opus 5 on TerminalBench 2.1. Analytics India Magazine separately said Fusion Max improves task-completion rates over top-tier frontier baselines while capping estimated inference cost at 1.1 times the baseline model. That 1.1x cap is not the same claim as the half-cost Claude Code result.
Fusion bills each model at its standard rate. Paid plans get a limited-time 30% discount on those costs. It is available on all Copilot plans in Visual Studio Code, the Copilot app, Copilot CLI, and JetBrains IDEs. Preview users of auto, HydraFusion, or HydraFusion Max move to the matching effort level.
Previously, developers would often choose a more powerful model for a complex task and forget to switch back for everyday activities.
Hari Pokhrel, principal software engineer at AT&T, said that in a statement GitHub gave The New Stack, and credited Fusion Standard with better token efficiency at AT&T.
Copilot App, Projects, and Enterprise Access
Jay Parikh, executive vice president of CoreAI at Microsoft, described an agentic system aimed at velocity, intelligence, and quality. The GitHub Copilot app covers specs, design discussion, and telemetry, not only code.
Copilot projects, in the app from October 28 and marked a public preview, run parallel agent sessions across repositories. Users attach a goal, repos, design mocks, or Slack threads and review pull requests. One cloud session continues across the app, VS Code, the CLI, and the web. A macOS and Windows computer-use preview asks for approval before controlling an app, and organizations can disable it.
Copilot code review Max posted the highest F1 score on ReviewBench for high-severity issues, Parikh wrote. Lite and Balanced stay for everyday reviews. Automations, triggered manually, by an event, or on a schedule, can propose improvements from past outcomes for someone to approve.
Enterprise customers can use these Copilot features without a separate subscription and pay for AI credits, through GitHub Sales now and self-serve soon. Copilot Business and Copilot Enterprise customers can keep current plans or move later.
Capacity Under the Agent Workload
Parikh wrote that agents are changing “the volume and shape of software development.” So far in 2026 GitHub has added 4 million CPU cores, 160 PB of high-speed storage, and 50 Tb/s of network capacity. An Azure multi-region migration is already underway and expected to be completed in early 2027. New Git architecture is meant to simplify primitives, decouple storage from compute, and minimize coordination.
Analytics India Magazine reported 7.38 billion commits in September, a 5x increase over the previous month, plus more than 3 million cores and 120 PB provisioned in August after reliability incidents. Those August figures are separate from GitHub’s year-to-date totals. The magazine also described AI code scanning on pull requests, reachability analysis with a call path, and an npm package firewall for known malicious packages, typosquats, and zero-day dependency attacks.
The New Stack noted Ramp moving model selection into Open SWE this month and Replit making intelligent routing the default over the summer. Compare Google’s response to the OpenClaw surge and Claude Mythos 5’s attacks on GitHub maintainers before widening unattended agent control.
