AI Pulse · AI Trends Pulse

GitHub is testing a different way to get better results from AI coding tools. Instead of sending every job to one large model, its research preview, Project HydraFusion, chooses between three approaches for each task. It can use a single model, run a cascade that checks quality before escalating, or have one model critique and revise another model’s work.

On Terminal-Bench 2.1, GitHub says this approach beat Claude Opus 5 by 4.9 percentage points on verified task quality, while producing an estimated cost 67% lower. That does not mean one model is suddenly obsolete. It does show that the best answer may come from deciding how work should be handled, not simply buying access to the most expensive model for every request.

For a business owner, the practical lesson is simple. Don’t treat AI model selection as a once-a-year software decision. Different jobs need different levels of speed, cost control, checking, and reasoning. A quick internal summary may need a cheap, fast model. A customer-facing proposal, financial analysis, or code change may justify a review step before anyone acts on it.

This is the kind of thing we build into an AI command centre, where work can be routed, checked, and monitored based on its business risk rather than handled with a one-size-fits-all AI setup.