GitHub Copilot's new HydraFusion preview does not ask you to pick a model. It picks a workflow - then runs whichever combination of models that workflow calls for - and in one benchmark it matched or beat Claude Opus 5 while cutting estimated inference cost by 67%. That number deserves a second read, because it reframes the whole question of how coding tools compete.
What GitHub Copilot multi-model orchestration actually does
Instead of asking developers to choose one model and accept its cost, latency, and capability profile for the entire task, HydraFusion creates an execution plan at runtime and can route work across models from multiple providers. That is different from what Copilot's earlier Auto model selection did. Auto selects one model per request. HydraFusion runs multiple models in a single turn, has them review and critique one another, and fuses the result. Auto picks the optimal model; HydraFusion picks the optimal workflow, possibly using multiple models.
The system currently selects among three workflow patterns: a single-model path, a cascade in which an efficient model drafts and a quality gate decides whether to escalate, and a critique workflow in which a second model family reviews the first draft before one revision.
The lineage matters for understanding how fast this moved. Auto V1 launched in January 2026 with capacity- and SKU-aware per-request selection, Auto V2 (HyDRA) followed in May 2026 with intent-scored routing, and HydraFusion arrived in August 2026 adding per-turn orchestration and cache-aware workflows. Nine months from "pick one model" to "build a bespoke pipeline per request."
The benchmark numbers are mixed, and that is the honest part
GitHub evaluated HydraFusion against Claude Opus 5 and GPT-5.6 Sol on three agentic coding benchmarks.
| Benchmark | vs. Opus 5 quality | Estimated cost vs. Opus 5 |
|---|---|---|
| TerminalBench 2.1 | +4.9 points | −67% |
| CheckpointBench (internal) | slightly below | −65% |
| DeepSWE | slightly below | −36% |
The reported figures are relative to Opus 5. CheckpointBench is GitHub's internal multi-turn set, curated from real Copilot sessions and anchored to immutable public commits so runs are replayable.
The published numbers do not imply that orchestration automatically dominates a frontier model. DeepSWE is the clearest counterexample: HydraFusion saved cost but gave up quality. The more defensible conclusion is that routing can move the quality-cost frontier when the policy correctly identifies which tasks need stronger inference, review, or escalation, and which tasks do not.
That caveat is the thing most coverage skips. A cascade or critique workflow only beats a single frontier call if the router knows when to escalate. When it misclassifies a hard task as easy, it ships worse output at lower cost - which is not a win.
Why this shifts how agentic coding tools compete
HydraFusion moves competition in AI coding from choosing one frontier model to controlling the router that allocates tasks, inference spending, and provider traffic. That is a structural change. Today, if you find Opus 5 too expensive for your team's median PR review, your option is to switch to a cheaper model and accept worse results on hard tasks. HydraFusion offers a third path: spend frontier-model money only on the slice of tasks that actually require it.
If you're picking between Claude Code, Codex, GitHub Copilot, and Cursor in late 2026, orchestration is becoming a real axis of comparison, not just raw model quality. Each platform is already diverging here. Claude Code exposes model and effort tradeoffs and subagent delegation as explicit, user-controlled choices. You decide when to delegate and to what. HydraFusion makes that decision automatically. One approach preserves developer control; the other optimizes without asking.
Billing is per token consumed by whichever models the workflow invokes, at each model's standard rate. There is no orchestration surcharge layered on top - yet. That pricing simplicity is worth noting, because it means a critique workflow that calls two models costs roughly what those two model calls would cost individually. For a team running a lot of code review, that math should be concrete: how many of your daily Copilot tasks are "hard enough to justify a critique pass" versus "fast enough to route to a cheaper single model"?
Beagle in action#engineering, 10:22amHow to actually try HydraFusion today
HydraFusion is live as a research preview for users on all GitHub Copilot plans, inside GitHub Copilot CLI only. There are no open weights and no self-hosted path.
To turn it on:
run /update, then /experimental on, then /model and select HydraFusion (Research Preview).
The GitHub Copilot app and VS Code are both targeting September as a fast follow.
This preview is designed to learn which tasks benefit from compound workflows and how orchestration affects latency and cost in practice. GitHub is explicitly collecting signal on where the router fails. If you use it, the session logs matter - not just to you, but to how the routing policy gets refined.
One practical thing to watch: latency. A critique workflow that calls two model families sequentially will be slower than a single fast model call. For autocomplete-style interactions that need a response in under two seconds, HydraFusion's cascade and critique paths will likely lose to a direct call. The benchmark results are on agentic, multi-turn tasks - not on inline suggestions. Keep that scope in mind when evaluating whether to enable it for your team.
GitHub Copilot multi-model orchestration: common questions
What is GitHub Copilot HydraFusion?
HydraFusion is a research preview inside GitHub Copilot CLI that orchestrates multiple AI models per coding task instead of sending every request to one fixed model. It selects from three workflow patterns - single-model, cascade, or critique - at runtime, based on the task's estimated complexity and the models' capability signals.
Does HydraFusion cost more than regular Copilot?
There is no orchestration surcharge. You pay the standard per-token rate for each model the workflow invokes. A critique workflow that calls two models costs roughly the sum of both model calls. On agentic benchmarks, GitHub's evaluation showed net cost reductions of 36-67% compared to routing everything through Claude Opus 5.
How is HydraFusion different from Copilot's Auto model selection?
Auto model selection picks one model per request. HydraFusion picks a workflow - potentially involving multiple models in sequence, with one reviewing or critiquing another - per request. The decision is no longer just which model to use; it is which sequence of model calls should solve the task.
Is HydraFusion available in VS Code or only in the CLI?
As of September 4, 2026, HydraFusion is CLI-only. GitHub has said the Copilot app and VS Code extension are targeted for a September follow-on release. Check /experimental in Copilot CLI to enable it now on any paid Copilot plan.
What is the real risk with multi-model routing for coding agents?
The router's classification accuracy is now the quality-determining variable, not any individual model's capability ceiling. If the router misclassifies a hard task as simple, it routes to a cheaper model and ships worse output. Teams should monitor task-type error rates - not just aggregate cost savings - when evaluating whether routing is actually working for their workload.