GitHub has put a research preview called Project HydraFusion into the Copilot CLI, and the idea underneath it is more interesting than the name. Instead of sending every coding request to one fixed model, HydraFusion orchestrates several, choosing between three execution patterns depending on what you have actually asked it to do. A quick autocomplete and a sprawling multi-file refactor are different jobs; HydraFusion stops pretending they should be handled by the same engine.
This is the direction the whole coding-agent field has been drifting for a year, and it is worth understanding even if you never touch the Copilot CLI. The competitive fight is moving up a layer, away from “whose model is smartest” and towards “who routes work to the right model fastest and cheapest.” For anyone paying per token, that routing layer is where actual money is won or lost.
The three patterns, in plain English
HydraFusion launched on 4 September as a research preview, available on every Copilot plan through the /experimental command in the CLI. When a request comes in, it picks one of three ways to handle it. Single is the obvious one: a single chosen model solves the task directly, used when the job is simple enough not to warrant anything fancier. Cascade is the thrifty one: a cheaper, faster model drafts an answer, then a quality gate decides whether that answer is good enough or whether the job should be escalated to a stronger, pricier model. Critique is the careful one: one model drafts a result, a separate read-only critic from a different model family reviews it, and the original model gets one shot to revise based on that feedback.
The clever part is that you do not choose. You select HydraFusion the way you would select any other model, and it works out, per request, which pattern balances quality, cost and latency best. GitHub frames workflow selection as an optimisation problem, reading capability signals for reasoning, code generation, debugging and tool use, then routing accordingly. You ask; it decides how much firepower the question deserves.
The numbers GitHub is leading with
Here is the figure doing the heavy lifting in every write-up, and it is worth stating precisely. On TerminalBench 2.1, one of three agentic coding benchmarks GitHub tested, HydraFusion improved verified task quality by 4.9 percentage points while costing 67% less than running Claude Opus 5 directly. Better work, at roughly a third of the price. If that holds up outside GitHub’s own evaluations, it is a genuinely awkward result for anyone selling a single premium model as the answer to every coding task.
Two caveats keep this honest. First, these are GitHub’s own offline evaluations, and the cost savings are described as “estimated”, so treat the exact percentages as a vendor’s best case rather than an independent finding. Second, HydraFusion is routing among the models GitHub has chosen to include; it is orchestration within a walled garden, not a neutral referee picking the best tool on the open market. The intelligence is real, but so is the incentive to keep you inside Copilot.
Why this matters beyond one CLI tool
Strip away the branding and HydraFusion is an admission that the “one giant model for everything” era is ending on economics alone. Frontier models are extraordinary and extraordinarily expensive, and most coding tasks do not need a frontier model; they need a competent one that answers quickly and cheaply, with the big gun held in reserve for the hard problems. Routing is how you get the quality of the expensive model at something closer to the price of the cheap one, and every serious player is now building this layer. Microsoft already swapped the brain inside Copilot once; this is the same instinct taken to its logical end.
For working developers, the practical takeaway is simple: your bill is about to depend less on which model you pick and more on how well the router picks for you. For everyone watching the industry, it is a signal that the value is migrating from the models themselves to the orchestration around them. The smartest model in the room stops mattering quite so much once something else decides when to actually call it.
Related on Top Tool Stack: Microsoft swapped the brain inside GitHub Copilot for its own and Cursor is a SpaceX company now.