GitHub Launches Copilot Fusion, an AI Model-Orchestration Engine for Developers

At the GitHub Universe conference on October 28, GitHub announced Copilot Fusion, a system that automatically selects and orchestrates multiple AI models to handle coding tasks such as reasoning, code generation, debugging, and tool use. Developers enable a single mode—Standard, High, or Max—and the system routes requests to the most suitable model(s) for each step.

The New StackVentureBeat

Copilot Fusion builds on two prior GitHub initiatives: Project HydraFusion, a research preview that demonstrated runtime model orchestration, and Copilot Auto Model Selection, which automated the choice of a single model. The technology lets the system operate in three execution patterns—Single, Cascade, and Critique—to either use one model directly, start with a cheaper model and escalate if needed, or have one model draft code while another reviews it. Enterprise administrators can restrict which models Fusion may select, preserving governance controls.

The New StackVentureBeat

• Benchmark cost: Artificial Analysis measured an average task cost of $5.59 for Fusion Max versus $11.42 for Claude Code with Anthropic’s Opus 5.5 Max, a reduction of roughly 51 %.

VentureBeat

• Performance score: Fusion Max achieved a composite score of 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code’s ?62.

VentureBeat

• Token usage: Fusion used about 7.2 million tokens per task, roughly half the token count of the Opus 5.5 configuration and about 30 % less than the competing Devin Fusion system.

VentureBeat

• Time savings: GitHub reports Fusion Max completed benchmark tasks in about 40 % less time than the Claude Code baseline.

VentureBeat

Product Officer Mario Rodriguez told VentureBeat, “It’s not a model. It’s an orchestration engine at runtime,” emphasizing that the system decides which models to invoke rather than being a model itself. He added, “We have a single mode where we just go directly into the model that we believe is the best for the task.”

VentureBeat

GitHub also highlighted enterprise-level controls: administrators can view and disable specific models for Fusion, set default effort levels, and receive visibility into model usage. The company announced a temporary 30 % discount on model costs for paid Copilot subscribers using Fusion.

The New Stack

• Cost-reduction claims: Microsoft CEO Satya Nadella cited potential savings of up to 77 %, but that figure comes from a separate internal evaluation, not the 51 % reduction measured by Artificial Analysis.

VentureBeat

• Completion-time data: The 40 % faster execution claim originates from GitHub’s own reporting; independent testing has not published the underlying per-task timing data.

VentureBeat

GitHub is rolling out a new AI-usage billing model for eligible Enterprise Cloud customers, moving from seat-based AI credits to metered usage that begins with the first request. Copilot Standalone will be priced at $10 per user per month for enterprise-account users without a GitHub Enterprise license.

VentureBeat

The company also announced infrastructure upgrades—including 160 petabytes of high-speed storage, 4 million CPU cores, and 50 terabits per second of network capacity added in 2026—and a shift to a multi-region Azure deployment slated for early 2027. These expansions aim to support the increased load generated by AI-driven development sessions.

VentureBeat