Model Fusion

Send one prompt to several models at once, and optionally merge them into a single answer.

POST/v1/fusion

Model Fusion is a Pleum-native endpoint. It does not follow the OpenAI or Anthropic SDK shape, so call /v1/fusion directly. It fans one prompt out to 2–4 models in parallel, and when fuse is true it synthesizes the successful answers into one merged answer. Authenticate by passing your plm_ key as a Bearer token.

Request body#

Parameter	Type	Required	Description
models	string[]	Required	Array of 2–4 model IDs. Duplicates are removed; fewer than 2 distinct models returns `400`.
messages	array	Required	Same `{role, content}` array as chat. Multimodal input is supported.
temperature	number	Optional	Default 0.7.
max_tokens	integer	Optional	Default 2048.
fuse	boolean	Optional	If `true` and at least 2 sources succeed, synthesizes a merged answer. Default `true`.
fuse_model	string \| null	Optional	Model ID used for synthesis. Default `null`, which uses the first successful source's model.

request

curl https://apirouter.pleum.ai/v1/fusion \
  -H "Authorization: Bearer $PLEUM_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "models": ["gpt-4o", "claude-sonnet-4-6"],
    "messages": [
      {"role": "user", "content": "Explain the CAP theorem in one paragraph."}
    ],
    "temperature": 0.7,
    "max_tokens": 2048,
    "fuse": true
  }'

HTTP (Python)

import requests

resp = requests.post(
    "https://apirouter.pleum.ai/v1/fusion",
    headers={
        "Authorization": "Bearer plm_...",
        "Content-Type": "application/json",
    },
    json={
        "models": ["gpt-4o", "claude-sonnet-4-6"],
        "messages": [{"role": "user", "content": "Explain the CAP theorem in one paragraph."}],
        "temperature": 0.7,
        "max_tokens": 2048,
        "fuse": True,
    },
)
print(resp.json()["text"])

HTTP (TypeScript)

const res = await fetch("https://apirouter.pleum.ai/v1/fusion", {
  method: "POST",
  headers: {
    Authorization: "Bearer plm_...",
    "Content-Type": "application/json",
  },
  body: JSON.stringify({
    models: ["gpt-4o", "claude-sonnet-4-6"],
    messages: [{ role: "user", content: "Explain the CAP theorem in one paragraph." }],
    temperature: 0.7,
    max_tokens: 2048,
    fuse: true,
  }),
});

const data = await res.json();
console.log(data.text);

Response#

sources holds each model's result (model, provider, text, cost_krw, latency_ms, error). When synthesis runs, fused_text and the model used for it (fused_by) are returned, and total_cost_krw is the sum of every source call plus the fuse step.

200 OK

{
  "sources": [
    {
      "model": "gpt-4o",
      "provider": "openai",
      "text": "The CAP theorem states that a distributed system...",
      "cost_krw": 3,
      "latency_ms": 912,
      "error": null
    },
    {
      "model": "claude-sonnet-4-6",
      "provider": "anthropic",
      "text": "In any distributed data store, you can guarantee...",
      "cost_krw": 4,
      "latency_ms": 1180,
      "error": null
    }
  ],
  "fused_text": "The CAP theorem says a distributed system can provide at most two of consistency, availability, and partition tolerance...",
  "fused_by": "gpt-4o",
  "total_cost_krw": 9
}

Each source call bills real credits, and the fuse step bills again on its own — total_cost_krw is the sum of all of them. Partial failures are isolated: a failed source returns with its error set, and the call still returns 200. The fuse step runs only when fuse is true and at least 2 sources succeed.