THE SIGNAL IN ONE SENTENCE
Instead of asking one AI model to solve every coding problem, GitHub can now construct a small workflow of models that draft, review, revise, or escalate the work.
01
WHAT ACTUALLY CHANGED
GitHub released Project HydraFusion as a research preview inside Copilot CLI. A developer selects HydraFusion as if it were one model, but the system builds an execution plan for the request using models from several providers.
HydraFusion currently chooses among three patterns. Single sends the job to one selected model. Cascade lets an efficient model try first, then uses a quality gate to decide whether a stronger model should take over. Critique lets one model draft, sends the result to an independent read-only critic from another model family, and gives the original model one revision pass.
The runtime accounts for drafting, critique, revision, escalation, retries, and fallbacks. Critics cannot use tools or modify the repository. If the workflow is cancelled or fails validation, GitHub says it applies no patch at all.
In GitHub's controlled evaluations, its best HydraFusion configuration scored 4.9 percentage points above Claude Opus 5 on TerminalBench 2.1 while using 67 percent less estimated cost. On DeepSWE it finished 1.5 points behind at 36 percent lower cost, and on GitHub's internal CheckpointBench it finished 0.1 points behind at 65 percent lower cost.
02
WHY THIS MATTERS
The familiar model picker assumes there is one best brain for a task. HydraFusion treats the decision more like computing infrastructure. A small request may need one quick model. A difficult repair may justify escalation. A subtle change may benefit more from disagreement than from another confident first draft.
Developers already perform this orchestration manually by asking one system to build and another to review. Automating that habit could improve quality without charging frontier-model prices for every step. It also creates a new layer of behavior that teams will need to evaluate. The router, critic, acceptance gate, and fallback policy can matter as much as any model inside the pool.
The most consequential part may be invisible. A person sees one answer and one patch while several models made decisions behind it. That makes complete cost accounting and traceable workflow choices essential, particularly when the result reaches production code.
03
WHERE IT COULD HELP
- Route routine repository work to inexpensive models
- Escalate difficult fixes only when a quality gate fails
- Use an isolated second model to review sensitive changes
- Audit the cost, latency, and outcome of every model call
KEEP A HAND ON THE WHEEL
HydraFusion is an active research preview. GitHub says first-turn, well-scoped coding tasks are the strongest fit today, while longer conversations still need work. Its benchmark results are company-run and depend on the evaluated models, prices, harnesses, and routing configuration.
04
TERMS WORTH KEEPING
OPEN GLOSSARY CARD
Tool calling
A model requests a defined action instead of trying to answer from memory.
OPEN GLOSSARY CARD
Agent orchestration
Coordinating several agents, tools, or specialist roles across one larger task.
OPEN GLOSSARY CARD
Evaluation gate
A required test or review that a system must pass before it advances to the next stage.
SOURCES AND VERIFICATION STATUS
This article was written from the materials below. Product claims and dates were checked against those sources on September 5, 2026.
PUBLICATION RECEIPT: Revision 1. Approved by Zak and published September 5, 2026.
THE PUBLICATION ENGINE
WANT A SIGNAL OF YOUR OWN?
We build source-grounded publications, private briefings, and editorial systems for organizations with something useful to say.
WORK WITH US