GitHub has shipped a model router called HydraFusion as a research release in the Copilot CLI. Rather than sending every coding request to one pre-selected model, it decides per request how the work should be done: one model alone, a cheap model that escalates when its answer fails a quality check, or a draft reviewed by a model from a different family. GitHub describes the result as frontier-level quality. The company's own benchmark table supports that claim in one test out of three. Cost dropped in all three, by as much as 67% against running Claude Opus 5 alone.
The distinction GitHub is drawing is worth getting right, because it is the whole product. Mario Rodriguez, GitHub's chief product officer, said conventional routing answers the question of which model will handle a task better. HydraFusion tries to work out how the task is best solved: whether one model is enough, whether to start with a fast one and move to a more powerful one if needed, or whether to bring in an independent model to check and revise the result.
Before the first model call, HydraFusion evaluates the request and picks one of three execution paths.
Single: one model solves the task immediately, when no extra checking or handoff is needed.
Cascade: a more efficient model prepares a solution first, then a quality check either accepts it or passes the same task to a more powerful model.
Critique: one model produces a draft, an independent model from another family reviews it in an isolated environment without tools, and the original model then makes one pass at fixing the result based on the notes.
The name points at its lineage. HyDRA — a hybrid dynamic routing architecture — is a routing paper Microsoft researchers published earlier this year. GitHub is productising its parent company's research inside the CLI its customers already run.
The numbers come from offline evaluations, with no internet access, across three coding benchmarks, against Claude Opus 5 and GPT-5.6 Sol. On TerminalBench 2.1, HydraFusion scored 4.9 percentage points higher on verified quality than Opus 5, at 67% lower estimated cost. On DeepSWE it came in 1.5 points below Opus 5, with costs 36% lower. On CheckpointBench it was 0.1 points below Opus 5, at 65% lower estimated cost.
Read that table as a table rather than as a press release and the shape is clear: three benchmarks, cost down in every one, quality matching or beating Opus 5 in exactly one. The 0.1-point gap on CheckpointBench is a rounding error and most buyers would take it for a 65% saving. The 1.5 points on DeepSWE is not, and it is the test where the saving is smallest.
Where the saving comes from is not mysterious, and a technical breakdown published online spells it out: the result is a distribution effect, not a capability gain. Developer Awan Farz noted on X that the cheap model runs on every request that goes through the cascade, while the expensive model is invoked only for the share of tasks that fail the quality check. The economics are set by that share, which means the interesting variable in the system is not any model in it but the accuracy of the checker deciding when to escalate.
Not everyone read the quality gap as a caveat. AI commentator Martin Szerment wrote on X that model choice has stopped being a standalone decision and become an implementation detail, and that the release shows per-task model selection turning into an infrastructure function.
GitHub has been here before. It launched a routing feature called Auto earlier this year. Rodriguez said the two operate at different levels: Auto determines which single model best fits a given task, while HydraFusion selects the combination of models and the sequence of steps that should produce the best result. In practice Auto picks a model and HydraFusion orchestrates a workflow — it can decide one model is enough, have one model draft while another independently reviews, or hand the task up to a more powerful model when the first attempt misses the quality bar. GitHub calls the two complementary and is exploring combining them.
The gap between marketing claims about quality and what the benchmark tables show is not a GitHub habit. Nvidia released NeMo Switchyard in August alongside Nemotron 3.5 Lightning, saying the system preserves frontier-level accuracy while cutting task cost to roughly a third of running Claude Opus 4.8 alone. The most detailed external benchmark Nvidia published was prepared by LangChain: across 145 multi-turn tasks, routing just 7% of requests to the frontier model cut costs by 74% and also produced a noticeable accuracy loss against frontier-only. OpenRouter shows the same pattern in its own figures. Its new Auto router, also launched in August, is claimed to beat the previous version across a broad range of tasks and cost levels; the published table confirms that in three of five test categories and shows the new router losing in the other two — MMLU Pro at 85.2% against 86.6%, and τ³-bench Banking at 20.6% against 21.0%.
My read is that "frontier-level quality" has become the routing category's version of a cost claim in disguise. Every one of these systems is really selling the same thing — a controlled, usually small quality sacrifice in exchange for a large and reliable cost reduction — and none of them wants to say it in those words, because the buyer's approval chain is easier to clear on quality than on savings. What is notably absent from GitHub's announcement is any statement of how much quality it considers acceptable to give up, or on which kinds of tasks. The table gives three data points; the policy behind the escalation threshold, which is the actual product, is not described at all.
The scope limit matters more than the benchmark spread. HydraFusion currently handles only first-turn coding tasks, one request at a time; GitHub says multi-turn coordination is still in development. A wrong escalation decision on a single turn costs one weaker answer. In a long agent session it seeds the context every subsequent turn builds on, and the cheap-model share that makes the economics work is exactly what raises the odds of that happening. GitHub is shipping the easy half of the problem first and calling the result frontier-level.
That is the tension buyers are now inside. Vendors are optimising two numbers and publishing one, and the routing layer is where the trade between them gets made — silently, per request, inside a tool engineering teams evaluate on quality alone.