GitHub's HydraFusion cuts cost. Read the quality column.

GitHub shipped a research preview called Project HydraFusion on September 4. Instead of sending every coding task to one model, it picks an execution strategy per task — one model, an escalation chain, or a draft-and-critique pair. The model routing results are the interesting part, and not for the reason the headline suggests: cost dropped in all three benchmarks, but quality only went up in one of them.

What actually happened

Per the GitHub blog post, HydraFusion chooses between three patterns at runtime. Single sends the task to one model. Cascade has a cheaper model draft, then a quality gate either accepts the draft or escalates to a stronger model. Critique has one model draft, a critic from a different model family review it, and the original model revise once.

The benchmark table, measured against Claude Opus 5:

• TerminalBench 2.1 — 67% lower cost, +4.9 points of verified task quality
• DeepSWE — 36% lower cost, −1.5 points
• CheckpointBench — 65% lower cost, −0.1 points

So one win, one rounding error, one real regression. VentureBeat put it plainly in its headline. GitHub CPO Mario Rodriguez frames the system as constructing an execution strategy rather than prompting a model. It is live now in the Copilot CLI on all Copilot plans: run /update, then /experimental on, then pick HydraFusion under /model.

Why model routing matters for your business

Two-thirds off your coding-agent bill is not a rounding error, and this is the first time a major vendor has published the cost and the quality delta side by side instead of only the win. Take the honesty and copy the method.

The pattern generalizes past code. Most of what your automations do is easy — classify a ticket, extract five fields from an invoice, draft a reply that a human will edit anyway. A cheap model clears that work. The hard 10% is what needs the expensive model, and you only find out which is which at runtime. That is what a cascade is: a cheap first pass plus a gate that decides whether to escalate. You can build one this week with a confidence threshold and a second API call.

The catch is the gate. A cascade with a bad gate is not cheaper — it is a cheap wrong answer plus a full-price retry. Before you route anything, you need a scored eval set of your own tasks, because a benchmark that says "−1.5 points" is telling you the escalation logic missed some cases. On your workload it will miss different ones.

Route the easy work down. Measure what the gate lets through. Keep the expensive model for the tasks that earn it.

Sources: GitHub Blog — Project HydraFusion, VentureBeat.