GitHub just shipped something that quietly undermines a core assumption in AI development: that you pick one model and live with it. Project HydraFusion, now live as a research preview in Copilot CLI, builds an execution plan per request. Sometimes one model handles everything. Other times an efficient model drafts, a quality gate checks the work, and a stronger model takes over if the first attempt falls short. There’s also a critique pattern where a read-only reviewer from a different model family flags issues before a single revision pass.
The numbers are worth pausing on. On TerminalBench 2.1, HydraFusion delivered 4.9 quality points above Claude Opus 5 while cutting estimated cost by 67%. That’s the headline result. But the other benchmarks tell a more complicated story: on DeepSWE and CheckpointBench, HydraFusion actually trailed Opus 5 slightly, though it still reduced costs by 36% and 65% respectively. GitHub’s team ran these evaluations at medium reasoning across all models, using Opus 5 and GPT-5.6 Sol as fixed baselines.
Engineers will care about the guardrails as much as the benchmarks. Every leg of a workflow—drafting, critique, revision, escalation, retry, fallback—gets logged with role, outcome, cost, latency, and diagnostics. Critics run in tool-less contexts and can’t touch the repository. If a workflow is cancelled or fails validation, no patch gets applied. That’s the kind of fail-safe design that distinguishes a production tool from a research toy, and it’s refreshing to see GitHub treat multi-model orchestration as an engineering problem rather than a prompt-engineering trick.
There’s a catch: this isn’t open. No weights, no self-hosted path, and it only works inside Copilot CLI. You enable it with /update, /experimental on, then select HydraFusion from the /model menu. Billing stays per-token at each underlying model’s standard rate, so the cost savings come from routing efficiency, not discounted inference. The broader implication is clear—model selection is becoming workflow selection, and the ‘pick one model’ dropdown may soon feel as dated as hardcoding a single database index. Whether HydraFusion’s approach generalizes beyond CLI coding tasks remains an open question, but the architecture is compelling enough that other providers will likely chase it.
