GitHub has detailed Project HydraFusion, a Copilot research preview that coordinates AI models within a coding task. In its September 4 announcement, the company describes a system that decides how to tackle a request, potentially using different models to produce and assess a solution.

 

The release puts a practical question at the center of AI coding: can a carefully chosen sequence of model calls deliver useful software changes more economically than assigning every job to one powerful model?

 

The benchmark headline needs context

GitHub’s published evaluation compares tuned HydraFusion workflows with Claude Opus 5 across three benchmarks. On TerminalBench 2.1, estimated cost was 67% lower and verified task quality rose 4.9 percentage points. DeepSWE showed 36% lower cost with quality down 1.5 points. On GitHub’s internal CheckpointBench, cost fell 65% and quality was 0.1 points lower.

 

These are controlled offline results tied to particular configurations and pricing assumptions. They do not establish that every developer will save 67%, or that HydraFusion beats the comparison model across all coding workloads.

 

The distinction matters because a small quality difference can have very different consequences depending on the task. A developer fixing a minor interface issue may tolerate another review cycle. A team changing a payment calculation may place much more weight on correctness than on the price of the initial attempt.

 

Three ways to handle a coding request

GitHub’s official community announcement describes three execution patterns. Single uses one model directly. Cascade starts with an efficient model and escalates if the result fails a quality gate. Critique adds an independent reviewer from another model family, followed by a revision from the original drafting model.

 

The preview is available through Copilot CLI across Copilot plans. GitHub directs users to update the CLI, enable experimental features with /experimental on, and choose HydraFusion from /model. The company says users cannot currently select or exclude particular models from HydraFusion’s curated pool, whose composition can change.

 

There is no separate HydraFusion fee, according to the same announcement: a turn’s charge combines the model work performed in its phases. Intermediate drafts are withheld while review and revision take place, leaving users with a final reviewed result. GitHub says it is working on better visibility during execution.

 

How this differs from Auto selection

Copilot already has a feature that makes model choices. GitHub’s Auto selection documentation says it considers task complexity alongside model health and availability. It also describes routing at natural cache boundaries to avoid unnecessary costs from changing models mid-session.

That existing feature provides useful context. 

 

Choosing a model and coordinating several stages of work are different decisions. For a simple request, the additional stages may have little value. For a task that benefits from an independent check, a reviewer could catch an issue that the initial solver overlooked. Whether that happens consistently is something users need to measure rather than assume.

 

What the cost claims mean for users

GitHub’s billing reference explains that Copilot interactions consume input, output and cached tokens, with prices depending on the model. Usage is converted into AI credits, valued at one cent each. Included allowances vary by plan, while Business and Enterprise allowances are pooled at the billing-entity level.

 

For an engineering manager, the relevant comparison is therefore the total effort required to get an accepted change. That includes model usage, waiting time and human review. A cheaper first response is not necessarily a cheaper completed task if it creates more follow-up work.

 

A useful pilot would give HydraFusion and an existing workflow comparable bug fixes, documentation changes and test-writing tasks. Teams could then record whether the changes pass their checks, how much review they need and how long completion takes. Keeping those categories separate would help reveal whether orchestration is valuable for particular types of work rather than only on average.

 

Related coverage

HydraFusion’s wider significance is that the workflow is becoming a product decision in its own right. Developers are being asked to evaluate how an assistant organizes work, alongside the capabilities of the individual models it uses. 

 

The preview offers a concrete opportunity to test that idea, with the strongest conclusions still dependent on results in real repositories.