You lost Fable 5. Don't buy three models in a trenchcoat.

Hermes MoA: real, tunable, and self-graded

Nous shipped MoA presets as 'virtual models' in late June: reference models run first, then an aggregator writes the actual response and tool calls using their analyses as private context. The headline — about 8% over Opus 4.8 and 11% over GPT-5.5 — comes from the launch post, on an upcoming, unpublished benchmark. Worth noting: Hermes's own docs report the win as roughly 6 points, not 8% — a gap they don't reconcile.

Either way it's self-reported on a benchmark nobody else can run yet, so treat it as a hypothesis. To its credit, the docs are candid that MoA multiplies the number of model calls, and you can tune which models it uses. Our verdict: the most honest of the three, and the only one you self-host and tune — but its proof is a benchmark of one.

OpenRouter Fusion: a great escalation lane, a terrible default

Fusion is the slickest productization: one endpoint fans out to a panel of up to eight models, each with web search, and a judge synthesizes consensus, contradictions, and blind spots. On OpenRouter's own deep-research set, a Fable-5-plus-GPT-5.5 panel beat every individual model, and a budget panel came within a point of Fable 5 at half the cost.

The catch is in the fine print. OpenRouter says a Fusion call is often 2-3x longer than a standard one; one developer who rebuilt the pattern measured 7x slower and 4x the cost versus calling Opus directly. And the benchmark it aces has no coding domain, so it tells you little about your IDE. Our verdict: brilliant for the rare high-stakes question where you'd happily pay 3x for a fused second opinion. Use it as an escalation lane, never autopilot.

Sakana Fugu: the cautionary tale

Fugu is the most ambitious: not a hard-coded pipeline but a trained orchestrator that learns when to answer directly versus delegate to a swappable pool of frontier models. Sakana claims Fugu Ultra stands shoulder-to-shoulder with Fable 5 and Mythos, topping 10 of 11 benchmark rows.

Then people used it. Within a day, independent testers — Ethan Mollick among them — reported a sharp gap between the leaderboard and real work: routine tasks taking around half an hour, with output that was 'fine' but didn't match Fable in practice. Routing is proprietary, so you can't see which models ran, and the announcement is silent on how much orchestration multiplies token cost. Our verdict: the clearest illustration of the genre's risk — frontier benchmarks, sub-frontier reality, black-box billing. Watch it; don't bet on it yet.

The quiet contender: just use GLM-5.2

Here's the take the orchestration pitch needs you not to consider: maybe you don't need an orchestration layer at all. GLM-5.2 is open-weight, MIT-licensed, and lands within about a point of Opus 4.8 on mainstream coding benchmarks — at roughly a sixth of the cost. It's not magic: it trails Opus badly on ultra-long-horizon work. But notice the arithmetic.

Two of the three orchestration tools cost more than running Opus directly and add latency; GLM-5.2 costs a fraction and runs as a single fast call. For the everyday work that fills most days, one strong cheap model beats three expensive ones in a trenchcoat. Our verdict: the smart default for anyone who lost Fable 5 and wants 90% of the capability at a sixth of the cost, today.

The takeaway