model: "fusion"
Fusion
Several inexpensive models work on your request in parallel, and an aggregator combines their answers into one reply. The result holds up against frontier models and costs less.
Why several models beat one
Every model has blind spots, and different models have them in different places. When several work on the same request and one weighs their answers, what one model misses another usually catches.
Several models on every request
A panel of models works on the same request at the same time. Each brings its own training and catches different things.
Answers checked against each other
The aggregator compares the panel's answers, keeps what holds up, resolves where they disagree, and writes one reply.
Frontier quality, lower price
Together, inexpensive models reach the level of frontier models like Claude Opus 5.5, and still cost less.
10%
higher score than Claude Opus 5.5
DRACO
50%
lower cost per task than Claude Opus 5.5
DRACO
Good fit
- Hard questions where one model's blind spot is costly.
- Deep research, design reviews, architecture plans, code review.
Not a fit
- Latency-sensitive chat or autocomplete. Use Auto or a single fast model.
- High-volume, low-stakes requests where cost per call dominates.
Cost and score, side by side
Each benchmark plots cost per task against score, next to single models run on the same harness and the same tasks.
Fusionmodel: "fusion"
Harness: TBD · Questions: TBD · Runs per configuration: TBD · Cost: every panel and aggregator call, measured tokens × public USD prices
fusion
- Protocols
- Chat Completions · Anthropic Messages · Responses
- Streaming
- Yes, the aggregator's reply streams
- Tool calls
- Yes
- Panel
- A fixed default panel of models and one aggregator
- Context length
- TBD
- Models used
- Inexpensive models on the panel. The panel changes between versions; your console lists every panel and aggregator call.
Ask Fusion the hard one
Change two settings in the agent you already use and see what it costs on your own work.