Read the Whole HydraFusion Table: It Beats Opus 5 on One Benchmark of Three, at 36%-67% Lower Cost

GitHub shipped Project HydraFusion as a research preview on 2026-09-04. In one line: you stop picking a model and let it assemble an execution plan per task — a cheap model drafts, a gate decides whether to escalate, or a model from a different family reviews the draft.

The official post publishes the full benchmark table, and that table is more honest than the headline. Verified 2026-09-08.

1. The official numbers, against Opus 5

So only one of the three benchmarks shows a genuine quality win. The other two are “slightly behind, at a third to two thirds less cost”. GitHub’s own wording for DeepSWE is “within 1.5 percentage points of Opus 5 while reducing cost by 36%”, and for CheckpointBench “within 0.1 percentage points at 65% lower cost”. Within, not better.

The post also states the limits: these are controlled offline evaluations, specific to “the evaluated benchmark revisions, workflow configurations, model pool, and pricing assumptions”, with all models evaluated at the same medium reasoning level. These are not the numbers you will see on your own repository.

2. What it saves is the call mix, not the price list

Easy to misread. The post is explicit: HydraFusion is available on all GitHub Copilot plans through /experimental in Copilot CLI, and usage is based on the tokens consumed by the models it uses, priced at each model’s standard rate.

Which means:

• No dedicated discount. Choosing HydraFusion does not lower any unit price.
• The saving comes from not reaching for the most expensive model every time. A cheap draft that clears the gate is never escalated.
• Failed legs still cost money. The first of five stated operating principles is complete accounting: drafting, critique, revision, escalation, retry and fallback are all counted.

The consequence: if most of your tasks need escalation anyway, your bill will not drop. The published savings rest on the premise that many tasks never needed the strongest model.

3. Three execution patterns, three different problems

1. Single. One model solves the task directly — speed preserved where one pass is enough.
2. Cascade. An efficient model drafts; a quality gate accepts it or escalates to a stronger model.
3. Critique. One model drafts, an independent read-only critic from a different model family reviews, and the drafting model revises once.

Two design details matter: review steps run in isolated, tool-less contexts, so the reviewing model cannot touch your repository, and no patch is applied when a workflow is cancelled or fails validation — you are not left with a half-applied change.

4. What this means in practice

Available now:

• To try it: run /update for the latest Copilot CLI, then /experimental on, then /model and pick HydraFusion (Research Preview).
• GitHub suggests starting with substantial, well-scoped tasks you can hand over in a single prompt. Strong multi-turn performance is explicitly the next phase, not this one.

Not yet:

• You cannot watch it work. HydraFusion holds intermediate drafts and returns one coherent result, because those drafts may be revised or discarded and showing them live would make unfinished work look final. GitHub concedes that waiting without visibility is a real trade-off.
• Results will move. In their words, results, models, workflows, availability, names and product behaviour may all change during the preview.

5. Why it matters

The last year of AI coding competition has been about whose model is strongest. HydraFusion points elsewhere: same models, lower cost, through orchestration. GitHub frames it as moving from choosing the best model to dynamically constructing the best way to solve each task.

Worth watching: if orchestration really does cut more than half the cost at near-parity quality, then “is my subscription allowance enough” becomes a question about routing efficiency rather than model list prices. For anyone who keeps hitting a monthly cap, that matters more than a price cut.

How Copilot credits are counted: Copilot AI credits explained. Plan comparison on the GitHub Copilot tool page.

Source read directly on 2026-09-08: GitHub’s official blog post Project HydraFusion: Frontier quality via multi-model orchestration (GitHub Staff, 2026-09-04). All benchmark figures (67%/+4.9, 36%/−1.5, 65%/−0.1) are GitHub’s own controlled offline results; we have not reproduced them and cannot verify the evaluation environment. GitHub states the results apply only to the evaluated benchmark revisions and pricing assumptions.