Perfect

"model": "perfect" runs several best-in-classAIs in parallel for the category and has one model critically synthesize the results. It's the “pit the most expensive, strongest models against each other” mode — for jobs where accuracy matters more than cost.

curl https://apirouter.pleum.ai/v1/chat/completions \
  -H "Authorization: Bearer plm_..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "perfect",
    "messages": [
      {"role": "user", "content": "Refactor this function for thread safety and explain the trade-offs."}
    ],
    "stream": false
  }'

How it works#

The intent is classified and the top N benchmark models for that category (default 2, price-agnostic) are called as parallel workers. If 2+ succeed, a synthesizer model combines and validates them into the final answer; if exactly 1 succeeds, that result is returned without synthesis; if all fail, the request is rejected.

The worker pool is the top N from the same score table as Benchmark Smart. Code requests pit the best code models against each other; reasoning requests pit the best reasoning models.

Easy requests (casual chat, etc.) take a fast path — a single model call instead of the deep orchestration — keeping average cost near normal routing. Only code and reasoning categories get parallel+synthesis (operator-configurable).

Cost structure#

Cost is the sum of N worker calls + 1 synthesis call(3 calls by default). Each child call is billed normally in its own session — the parent just aggregates the results. If some workers fail, it falls back to a surviving worker's result and skips synthesis to save cost.

Related modes#

Benchmark Smart finishes in a single best-model call (faster, cheaper). To call a fixed set of models in parallel instead of auto-picked synthesis, use the Fusion endpoint.