Multi-model quality on a budget.
Stop waiting for the next model — the next jump in capability is a panel, not a checkpoint. Fusion sends your hardest prompts to several frontier models at once, then a judge weighs where they agree, catches where they contradict, and returns the single strongest answer — cross-checked against the rest. Reach for it on the work that actually matters — deep reasoning, hard coding, research, expert critique — anywhere being wrong costs more than a few extra completions.
Casual chat goes straight to a cheaper default. Coding, agent, and tool-using requests fan out to the panel — where a second opinion actually pays off.
Your prompt hits several frontier models at once — independent attempts that each catch what the others miss.
A judge reads every candidate and returns the single strongest one — cross-checked against the others, not a single model's first try.
Only where it helps. Fusion fans out to the panel on coding, agentic, tool-using and hard prompts — casual chat goes straight to a cheaper default, so you pay for the panel only on the requests where a second opinion pays off.
A single model gives you one attempt at your prompt — and no way to catch it when it is wrong. Fusion runs several frontier models on the same prompt at once, then an independent judge reads every answer and returns the strongest one, cross-checked against the others. Disagreement between strong models is signal: it surfaces the hard cases where one model alone would quietly get it wrong.
One model, one answer. No second opinion — if it is wrong, you ship it wrong.
Picks one model per prompt to save cost. Smarter selection — but still a single answer.
A panel answers in parallel; a judge returns the strongest, cross-checked against the rest.
| Model | Role | Context |
|---|---|---|
| Gemini 3.5 Flash | Panel | 1M |
| MiniMax M2.7 | Panel | 204K |
| GLM 5.1 | Panel | 200K |
| Claude Opus 4.8 | Judge | 1M |
Want a different combination? Build your own panel in the routing DSL editor.
Fusion is not special-cased — it is a built-in config of our routing DSL, a small sandboxed language for defining exactly how each request is routed. Clone the "Fable 5 Level" template and change the panel, the judge, the gating rule, or the way the winner is chosen.
A judge model ranks every candidate and serves the strongest one verbatim — never a diluted merge.
The panel votes on a structured answer — no extra judge call.
An aggregator model fuses the candidates into one improved answer (Mixture-of-Agents).
For code: the candidate whose patch actually passes your test harness wins.
A race — the lowest-latency leg wins, for when speed matters most.
No black box. Build your own router on the DSL and every fan-out is recorded — which models raced, each leg’s latency and status, and which answer the judge picked — so you can see in the console exactly why an answer won.
import os
from openai import OpenAI
client = OpenAI(
base_url="https://api.orcarouter.ai/v1",
api_key=os.environ["ORCAROUTER_API_KEY"],
)
response = client.chat.completions.create(
model="orcarouter/fusion-flash",
messages=[{"role": "user", "content": "Hello"}],
)
print(response.choices[0].message.content)Because fusion runs every panel member plus a judge, a fan-out request is billed as the sum of those underlying completions rather than a single model — and only on the requests that actually fan out. Zero markup.