GitHub has announced Project HydraFusion, a research preview in GitHub Copilot that automatically chooses how to solve each coding task — including which models to call and in what order — instead of relying on a single model. In GitHub’s controlled offline evaluations, HydraFusion’s selective workflows matched or exceeded the evaluated Claude Opus 5 baseline while reducing estimated workflow cost by up to 67%. The announcement is currently based on GitHub’s own official blog post, with no independent follow-up coverage in today’s news feeds.

How HydraFusion decides what to do

HydraFusion treats workflow selection as an optimization problem. For each request it reads capability signals for reasoning, code generation, debugging, and tool use, then picks one of three execution patterns:

  • Single. One selected model solves the task directly, preserving speed and efficiency.
  • Cascade. An efficient model drafts a solution and a quality gate decides whether to accept it or escalate to a stronger model.
  • Critique. One model drafts a result, an independent read-only critic from a different model family reviews it (following the same review pattern as Rubber Duck), and the drafting model revises once.

GitHub says the goal is to use additional model calls only when they are likely to improve the result. The runtime is built around five operating principles: complete accounting of cost and usage across every workflow leg, bounded execution with explicit timeouts, isolated review in tool-less contexts, fail-safe application that never applies a partial patch, and validated routing that checks model bindings and fallbacks before execution.

The benchmark results

GitHub evaluated fixed HydraFusion policies across three agentic coding benchmarks against Claude Opus 5 and GPT-5.6 Sol at the same medium reasoning level:

  • TerminalBench 2.1: 4.9 percentage points higher verified task quality at 67% lower estimated cost.
  • DeepSWE: within 1.5 points of Opus 5 at 36% lower estimated cost.
  • CheckpointBench (GitHub’s internal benchmark built from real Copilot sessions): within 0.1 points at 65% lower estimated cost.

GitHub is careful to call these controlled offline results specific to the evaluated benchmark revisions, workflows, model pool, and pricing assumptions — the research preview exists to find out how they translate to real developer workloads.

Why this matters for coding agents

The pattern behind HydraFusion — draft, review, escalate, and account for cost — is the same shape of workflow that makes agentic coding practical, whether it runs inside a hosted product or a local client. The difference is where the control lives. GitHub orchestrates models on its side and hides the complexity; a local approach like MOOGH keeps the orchestration visible and permissioned on your own machine, so you can combine local and cloud models while files and approvals stay with you.

Whichever way you work, the takeaway is the same: the next step in coding agents is less about any single model and more about reliably composing models into verifiable workflows. If you want that control on your own PC, download MOOGH for Windows and point it at the model you trust.