Here we are: routing and orchestration are converging on GitHub Copilot.
By “routing,” we mean the Auto mode. Introduced at the start of the year. It has taken two forms. Version 1 used a ModernBERT encoder to steer requests based on model availability and capabilities. Version 2, called HyDRA (Hybrid Dynamic Routing Architecture), uses the same type of classifier, but trained to assess needs across four dimensions: reasoning, code generation, debugging, and tool usage. It also incorporates operational constraints such as session continuity and dynamic failure filtering.
The orchestration component is newer. GitHub began experimenting with it in early September, under the name HydraFusion. The promise: for each prompt, to choose not the optimal model, but the optimal workflow, among three options:
- Task execution by a single model
- Intervention by a judging LLM that can validate or decide to entrust the task to another model
- An iteration loop with a critical model that only intervenes in reading
Local models remain in preview
From the convergence of the two components emerges the Fusion mode. It can operate at four levels of effort: Standard, High, Max, and Fast, the latter not adjustable by the user (it is reserved for the inline chat in VS Code).
The Standard level is accessible on all Copilot plans. The High and Max levels are available only on Pro+, Max, Business and Enterprise subscriptions. While Auto mode previously granted a 10% discount on model prices, GitHub is offering 30%… but for a limited time (through the end of November).
On the admin side, you can choose to enable Fusion by default in the picker. And you can control the model pool, as well as the default effort level for new conversations.
The integration of local models remains in preview. It currently concerns two:
- MAI Code 1.1 Flash, usable on Windows via the CLI and the GitHub Copilot or VS Code app
- Qwen3.8 Flash, usable on Windows, Mac and Linux via the CLI
Initial results were mixed at Michelin
Over the first few weeks of testing, frustration surfaced in particular around the lack of feedback during task execution. GitHub notably chose not to display intermediate step results. Rationale: to avoid customers perceiving them as finished work…
There was also a suggestion to GitHub to expose, rather than the chain of thought, structured telemetry. For example, the chosen execution strategy, the model families used, or patterns of result rejection.
Michelin was among the companies that tested HydraFusion. He spoke about it in mid-September. His assessment was mixed. In his view, this orchestrator was not a substitute for Auto mode, but a “quality mechanism”… at a clearly higher price.
Several user feedback responses echoed this sentiment. In broad terms: HydraFusion tends to pick overly capable models, and does not leverage “economical” sub-agents unless explicitly requested.
