GitHub's HydraFusion picks a model workflow per task in Copilot CLI
The research preview turns on in three commands and routes each request as single, cascade, or critique — real cost cuts, but it bills every model leg.
Copy markdown
Live now, three commands to switch it on
It's a research preview open to every GitHub Copilot plan, though CLI-only for now. Run /update, then /experimental on, then /model and pick HydraFusion (Research Preview) — no waitlist, no separate signup.
One router, three playbooks per task
For each request it reads capability signals — reasoning, code-gen, debugging, tool use — and picks the cheapest workflow likely to clear the bar: Single (one model), Cascade (a cheap model drafts, a quality gate escalates on failure), or Critique (a different model family reviews the draft, then the original revises once).
67% cheaper than Opus 5 — on one benchmark
GitHub clocks 67% lower cost than Claude Opus 5 on TerminalBench 2.1, with +4.9 points of verified quality. The catch: on DeepSWE it's 36% cheaper but −1.5 points, and on internal CheckpointBench 65% cheaper at −0.1. The cost win is real; the quality delta swings by task, so benchmark it on your own workload first.
It bills every leg — do the break-even math
You pay each model's standard token rate for every phase, so "the cost of a turn is the sum of its phases": Cascade is 1–2 calls, Critique is 3 by design. A cascade stops saving once escalation passes roughly 60% (Sonnet drafting) or 80% (Haiku) — and hard, multi-file tasks push escalation up, which is why DeepSWE saved less than TerminalBench.
Elsewhere: a free medical-reasoning model on Vercel
Inclusion AI's Ling 3.0 Flash Sante — a 124B Mixture-of-Experts (~5.1B active), 256K context — is free on Vercel AI Gateway through Oct 4 as inclusionai/ling-3.0-flash-sante-free. It's tuned for medical reasoning and evidence retrieval while keeping general coding ability.