GitHub, Spotify and CodeRabbit all decided the frontier model shouldn't be writing the code

GitHub published Project HydraFusion on September 4. Spotify published Portal on September 3. CodeRabbit published its Astra evaluation on September 4. None of them coordinated, and all three are the same argument.

HydraFusion is a Copilot research preview that treats workflow selection as an optimization problem. Each request gets routed to the cheapest of three patterns: one model solving directly, a cheap model drafting with escalation to a stronger one, or a draft reviewed by a critic from a different model family before revision. Against Claude Opus 5 and GPT-5.6 Sol baselines, GitHub reports 67% lower cost with +4.9 points on TerminalBench 2.1, 65% lower cost at -0.1 on CheckpointBench, and 36% lower cost at -1.5 on DeepSWE (GitHub Blog). Cheaper and better on the hardest of the three benchmarks. The cross-family critique step is doing the work there. It's the reviewer coming from a different vendor, not the extra pass.

Spotify's Portal is three layers and you could build it this afternoon. PreToolUse hooks that block file reads over a configurable threshold. Bash wrappers around a CLI that report token counts. A markdown skill telling Claude when to delegate. Bulk file reads and boilerplate go to cheap models on ephemeral runtimes, frontier models keep the reasoning. Mean bulk-read savings around 90%, against a stated baseline of engineering leaders spending $200 to $2,000+ per developer per month (Spotify Engineering).

CodeRabbit published the arithmetic that justifies not defaulting to the top model. On labeled bugs, Astra got 61.3% actionable coverage, GPT-5.6 Sol 59.0%, Opus 5 50.2%. On cross-file reviews the spread widens: 57.1%, 47.6%, 42.9%. At a fixed 100K input / 10K output task, Astra costs $1.50, Sol $0.60, Terra $0.32, Luna $0.032. So Astra is 2.5x Sol's price for 2.3 points of coverage, and 47x Luna's (CodeRabbit).

The same thesis showed up from the opposite direction in a 493-star MIT skill called fable-orchestrator, created September 2. Fable 5.1 plans and adjudicates and is barred from the implementation graph entirely. GPT-5.6 Luna implements. DeepSeek V4 Flash handles loops and high-throughput iteration (GitHub).

Now the pushback, because it's the most useful part. The Hacker News thread on Portal reached 156 points and the top criticism is that Spotify reported token savings with no quality measurement anywhere in the writeup. One commenter says the cheap worker model "found surface-level patterns but missed a subtle thread-safety bug" (Hacker News). That is precisely the failure delegation produces, and it's invisible in a token counter.

I'd take the PreToolUse hook and skip the celebration. Cap your bulk reads, route them to a cheap model, and then measure whether your review pass still catches the class of bug that costs you a weekend. A 90% cost cut that adds one production incident a quarter isn't a cost cut. GitHub measured quality alongside cost and published both directions of the delta, including where it lost points. Spotify published one number. Copy GitHub's disclosure, not Spotify's.