共有:
Project HydraFusion

Copilot Gets Smarter, Cheaper.

Until last month, GitHub Copilot CLI made you pick one model at the start of a session and stick with it the whole way through. HydraFusion, launched as a research preview on September 4, now auto-selects and combines multiple models per task. On the TerminalBench 2.1 benchmark, GitHub reports a higher score than Claude Opus 5 alone, at an estimated 67% lower cost.

AI Navigate Editorial·2026.09.08·6 min read
UNTIL LAST MONTH Task Opus 5 fixed pick same model for every task SINCE SEPT 4 Task route Single Cascade Critique picks 1 of 3 patterns per task
01
Before / After

The "fixed cost" of
picking a model is gone

Until now, GitHub Copilot CLI made you pick one model in the picker at the start of a session and stick with it for the whole task. That meant paying for an expensive top-tier model on a trivial fix, or the reverse: picking a cheap model for a hard task and eating the rework.

According to GitHub's official blog, "Project HydraFusion" shipped as a research preview on September 4, 2026, available to all Copilot plans through the `/experimental` flag in Copilot CLI. The same details were shared the same day in GitHub's Community Discussions announcement thread. HydraFusion builds a full execution plan per request and picks among models from multiple providers, choosing one of three patterns: Single (one model solves it directly), Cascade (a cheap model drafts and a quality gate decides pass/fail), or Critique (a separate model family reviews the draft and the original model revises once).

Before (fixed model)HydraFusion
Pick one model manually at session startAuto-builds an execution plan per task
Same model even for trivial fixesTrivial tasks go straight to a cheap model via Single
Rework on hard tasks can go unnoticedA failing quality gate escalates to a stronger model
Billed at the one model's rateBilled per model actually used, at each model's own rate

02
How It Works

Cascade reshapes cost around
"draft, then gate"

Of the three patterns, Cascade is the one that matters most in practice: a cheap model drafts, and only fails that miss the quality gate get escalated to a stronger model.

Draft cheap model quality gate pass shipped (1 call billed) fail escalated to a stronger model draft + full pass: 2 calls billed
FIG. A pass bills one draft-model call; a fail bills the draft plus a full escalated pass
01

Draft

A cheap model such as Haiku 4.5 or Sonnet 5 attempts the task first.

02

Quality gate

An automatic check decides whether the output clears the bar for the task. If it passes, the run is done.

03

Escalation (only on failure)

Only when the bar isn't met does a stronger model redo the task from scratch. Intermediate output isn't shown — only the final result comes back.

03
Benchmarks

What three benchmarks
actually showed

GitHub's own offline evaluation shows the results diverging depending on the benchmark.

-67%
TerminalBench 2.1 cost (quality +4.9pt)
-36%
DeepSWE cost (quality -1.5pt)
-65%
CheckpointBench cost (quality just short of Opus 5)

On TerminalBench 2.1 — an agentic benchmark of roughly 89 real terminal-operation tasks — GitHub reports HydraFusion's estimated cost running 67% below Claude Opus 5 alone, with accuracy 4.9 percentage points higher. That's the result behind today's item, "Copilot Gets Smarter, Cheaper Multi-Model Runs." But on the same published table, DeepSWE cost dropped 36% while quality actually fell 1.5 points below Opus 5, and CheckpointBench cost fell 65% while quality landed just short of Opus 5 as well. Only one of the three benchmarks — TerminalBench 2.1 — actually shows a quality win, a point that matters for the counter-argument below.

04
Who Benefits

Who this helps, and how

The same feature lands differently depending on your role.

Engineers

Tasks spanning multiple files, or with ambiguous requirements, are where this pays off most. That said, for a single quick one-line fix, escalation rarely triggers in the first place, so the gain is likely to be marginal. Since intermediate reasoning isn't shown — only the final result — that's worth keeping in mind while debugging.

PMs

One less decision — "which model" — for the team to make, freeing up time for scoping and prioritization. But quality still varies by task type, so "HydraFusion picked it, so it's fine" shouldn't become the release bar on its own; a review rule is still worth having.

Business / Decision-makers

No new fee — savings show up inside the existing token-based Copilot billing. Still, as the counter-view below shows, the actual outcome depends heavily on task mix; it's safer to look at a month of real usage before budgeting around the headline numbers.


The cost savings can vanish, depending on the escalation rate.


05
Risk / Counter-view

Why this isn't an unqualified win

While GitHub itself frames HydraFusion as delivering "frontier-level quality," a VentureBeat analysis points out that of the three benchmarks GitHub published, quality matched or beat Opus 5 in only one — TerminalBench 2.1. Cost fell across all three, but "faster, cheaper, and smarter, all at once" doesn't appear to hold uniformly.

The same article cites a more concrete analysis from developer Awan Farz: under the Cascade pattern, a pass costs one cheap-model call, but a fail costs the cheap-model draft plus a full pass from the stronger model — two calls. That means once the escalation rate crosses a certain break-even point, Cascade can end up costing more than sending the task straight to Opus 5. The analysis puts that break-even around a 60% escalation rate for a Sonnet 5 draft, and around 80% for a Haiku 4.5 draft. Teams whose workload skews toward harder tasks are the ones most likely to approach that threshold.

06
What To Do Next

What to do next

01

Try it via /experimental first

Copilot CLI is the only surface today. Run `/update` to get the latest build, then `/experimental on` and pick HydraFusion from `/model` — no extra fee to try it.

02

Measure it against your own workload

Published benchmark results vary by task mix. Track your escalation rate and cost over a week or two of real use before committing to it broadly.

03

Wait for VS Code / the Copilot app if you rely on them

GitHub has flagged VS Code and the Copilot app as fast-follow targets within September. Teams outside the CLI may want to time adoption around that rollout.