GitHub has launched Project HydraFusion as a research preview in GitHub Copilot, introducing runtime model orchestration for coding tasks.
Rather than relying on a single model throughout a task, HydraFusion selects and combines artificial intelligence models as needed. The approach lets Copilot draft, review and revise work, or shift a request to a stronger model when a simpler route does not meet the required standard.
The preview is available across GitHub Copilot plans through the command-line interface under an experimental setting. Users can select HydraFusion as they would any other model, while the underlying system determines how to handle each request.
The research builds on GitHub's existing automatic model selection system. In June, more than 9 billion requests were routed through its automatic mode, and more than half of paying Copilot users chose to let GitHub pick the model for their task.
How it works
The orchestration system treats workflow selection as an optimisation problem. It uses signals tied to reasoning, code generation, debugging and tool use to choose the least complex workflow expected to meet the quality threshold for a given task.
HydraFusion currently uses one of three patterns. In the single pattern, one selected model handles the task directly. In the cascade pattern, a lower-cost model produces an initial answer, and the system either accepts it or passes the work to a stronger model. In the critique pattern, one model drafts a result, a second model reviews it in a read-only context, and the drafting model revises the work once.
The review stage is isolated from the repository, so the critic model cannot alter code directly. By contrast, solver steps work within the shared workspace and follow the usual permission controls.
The system also records the role, outcome, cost, latency and diagnostics of each stage internally, while returning a single result and one change set to the developer. Intermediate drafts are not shown during execution because they may be revised or discarded later.
Cost and quality
GitHub published benchmark results from controlled offline evaluations across TerminalBench 2.1, DeepSWE and CheckpointBench, an internal benchmark based on real GitHub Copilot sessions. HydraFusion was compared with Claude Opus 5 and GPT-5.6 Sol under the same task inputs, tools, execution limits, pricing assumptions and grading conditions.
On TerminalBench 2.1, HydraFusion improved verified task quality by 4.9 percentage points while cutting estimated workflow cost by 67 per cent compared with Claude Opus 5. The benchmark measures coding agents on multi-step tasks in terminal environments.
On DeepSWE, which tests repository-level software engineering work across larger codebases, HydraFusion came within 1.5 percentage points of Opus 5 while reducing cost by 36 per cent. On CheckpointBench, it came within 0.1 percentage points of Opus 5 at 65 per cent lower cost.
All measured costs included each part of the workflow, including drafting, critique, revision, escalation, retry and fallback. The reported figures depend on the benchmark versions, workflow settings, model pool and pricing assumptions used in the evaluation.
Developer demand
The launch reflects a wider effort by GitHub to reduce the complexity developers face as the number of available models grows. Rather than requiring users to decide when to swap models or request a second opinion, HydraFusion attempts to make those decisions during execution.
Developers already manage this process manually by using one model for an initial answer, another for review, or a more advanced model for difficult problems. HydraFusion brings those steps into one system that can decide when additional model calls are likely to improve the result.
GitHub also outlined five operating principles behind the design: complete accounting of usage and cost across all workflow steps, bounded execution with timeouts and cancellation behaviour, isolated review, fail-safe application of code changes, and validated routing before execution begins.
Pricing in the preview is based on the tokens consumed by whichever models HydraFusion invokes, charged at each model's standard rate. That means usage costs can vary depending on the route the orchestration system chooses for a task.
GitHub said the first version is best suited to first-turn, single-prompt coding tasks. The research effort is still active, and results, models, workflows, availability, names and product behaviour may change as GitHub learns from the preview.
"You choose HydraFusion once and stay focused on your task while it manages the models and workflow behind the scenes," GitHub said.