GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 —
GROK-46H 865 —
GEM-37FH 865 —
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 —
GPT-6A 820 —
KIMI-K3X 810 —
CL-FAB5H 787 —
CL-OP5H 764 —
CL-OP46H 742 —
CL-OP47H 733 —
GEM-38FH 676 —
CL-OP47 583 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 —
GROK-46H 865 —
GEM-37FH 865 —
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 —
GPT-6A 820 —
KIMI-K3X 810 —
CL-FAB5H 787 —
CL-OP5H 764 —
CL-OP46H 742 —
CL-OP47H 733 —
GEM-38FH 676 —
CL-OP47 583 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
← Back to feed

GitHub HydraFusion Matches Opus 5 Quality on TerminalBench at 67% Lower Estimated Cost

GitHub has released HydraFusion into research preview inside GitHub Copilot. The system selects models from multiple providers at runtime — Claude, GPT, Gemini, and others — and routes each coding task through whichever execution pattern it judges most efficient.

The pitch: frontier quality without always paying frontier prices, by treating model selection and workflow orchestration as a single optimization problem rather than a static preference.

Three Patterns

HydraFusion picks from three execution modes per request:

  • Single: One model solves the task directly. Fast, cheap, appropriate when one model can handle it.
  • Cascade: An efficient model drafts; a quality gate decides whether to accept or escalate to a stronger model.
  • Critique: One model drafts; a read-only critic from a different model family reviews it (isolated from the repo, no tool access); the drafting model revises once.

The critique pattern borrows from GitHub’s existing Rubber Duck review agent. The cascade pattern is the cost-efficiency mechanism — most tasks never need to escalate.

Benchmark Results

GitHub evaluated HydraFusion against Claude Opus 5 and GPT-5.6 Sol baselines across three agentic coding benchmarks. All runs used the same task inputs, tools, execution limits, and medium reasoning level.

BenchmarkCost vs. Opus 5Quality vs. Opus 5
TerminalBench 2.1−67%+4.9 points
DeepSWE−36%−1.5 points
CheckpointBench−65%−0.1 points

CheckpointBench is GitHub’s internal benchmark built from real Copilot sessions. On two of three benchmarks, HydraFusion beats Opus 5 on quality while costing significantly less. The 1.5-point drop on DeepSWE at 36% lower cost is the only underperformance — modest enough that the routing economics still favour HydraFusion for most workloads.

The Engineering Constraints

GitHub built HydraFusion around five operating principles that matter for production use: complete cost accounting across every workflow leg, bounded execution with explicit timeouts, isolated review contexts (critic runs have no repo access), fail-safe patch application (no changes land if a workflow is cancelled), and validated routing that checks model availability before execution begins.

The developer-facing interface is intentionally simple: select HydraFusion like any other model in Copilot. The orchestration is invisible.

Implications

HydraFusion makes explicit what developers have been doing manually — picking different models for drafting versus reviewing, escalating hard problems to more capable systems. Automating that routing at the inference layer, across provider boundaries, with cost accounting per leg, is a meaningful step toward compound AI systems as a first-class product concept rather than a DIY architecture pattern.

As GitHub frames it: when new frontier models enter Copilot’s model pool, HydraFusion can evaluate and incorporate them without requiring user-facing changes. The routing absorbs capability improvements automatically.

The research preview is live now. GitHub says it will use production workload data to validate how the offline benchmark results translate to real developer usage.