GitHub's HydraFusion routes between models to match Opus 5 at 67% less cost

GitHub is shipping a new way of building coding agents: instead of picking one model and hoping it holds, its Copilot runtime now drafts, critiques, and escalates between models on the fly. The first benchmark numbers suggest the orchestration is doing real work, not just splitting the bill.

GitHub's Project HydraFusion matches Opus 5 on agentic coding at a fraction of the cost. The research preview, live in Copilot CLI today, treats workflow selection as an optimization problem. For each request, the runtime picks the cheapest of three execution patterns — direct, draft-and-revise, or cascade to a stronger model — using capability signals for reasoning, code generation, debugging, and tool use. On TerminalBench 2.1, HydraFusion improved verified task quality by 4.9 percentage points over Claude Opus 5 at an estimated 67% lower workflow cost; on CheckpointBench (an internal benchmark built from real Copilot sessions), it came within 0.1 points of Opus 5 at 65% lower cost; on DeepSWE, the gap was 1.5 points at 36% lower cost. The point isn't that any single model is best — it's that no single model needs to be. Developers select "HydraFusion" the same way they'd select a model, and the orchestration stays behind the curtain, with each leg's role, cost, and outcome logged for post-hoc review. The preview is intentionally scoped to first-turn, single-prompt coding tasks for now; multi-turn sessions are the next milestone. The bet is the same one Microsoft is making across the stack: the marginal frontier gain comes from how you stitch models together, not from any one of them.

Gerganov says llama.cpp stays hardware-agnostic under Nvidia-owned Hugging Face. Hours after Nvidia's $12.9 billion acquisition of Hugging Face closed, llama.cpp's lead maintainer Georgi Gerganov posted a public commitment that the project will "stick to its founding principles" and keep all backends — AMD, Apple Silicon, Intel, ARM, Nvidia — supported as usual. Nvidia engineers have been contributing to llama.cpp "for more than a year," he wrote, and that won't change. An Nvidia local-AI team lead, Gerardo Delgado, replied that the team's "goal is to grow Local AI" by providing tools that work "across IHVs" and that "llama.cpp stays neutral." For the open-weights community, the assurance matters: llama.cpp is the de facto runtime for running frontier models locally on consumer hardware, and any quiet tilt toward CUDA-only would have been a generational shift in the offline-AI ecosystem. The risk isn't gone — Nvidia now has both the most popular model hub and a significant voice in the most popular local-inference codebase — but the public pledge and the engineering reality (Apple Silicon, AMD ROCm, and Vulkan backends are all still under active development) make a near-term fork unlikely.

With frontier models closing in on cost parity through orchestration rather than raw scale, do you trust the runtime to pick the right model for the job — or do you want the choice back in your hands? Tell us in the comments.