Outperform frontier models with Merge Fusion

Introducing Merge Fusion, intelligence that beats frontier models at a quarter of the price.

Fusion sends one prompt to a panel of models, then uses a judge model to combine their answers into a single response that's more powerful than any one of them alone.

Why we built Fusion

Most AI products make one model do the whole job: handle the prompt and write the final answer. We found that the strongest agent systems split that work up. Several models work the problem independently, then one model reconciles what they found into a final answer.

Models disagree, and that disagreement is a useful signal. But even models that agree got there differently, calling different tools and weighing evidence differently along the way, and that difference carries its own insight. 

Fusion's judge folds in both. The result: responses that beat Fable-level answers from models that cost less than Fable.

How it works

A prompt fans out to every model in the panel in parallel, and each model answers the prompt independently. A judge model then synthesizes their responses and folds in its own analysis before writing the final answer.

You and your customers choose what models form the panel. Pick the analysis models and the synthesis model, and Fusion handles the rest:

curl --location 'https://api-gateway.merge.dev/v1/responses' \
--header 'Authorization: Bearer ' \
--header 'Content-Type: application/json' \
--data '{
    "model": "fusion",
    "input": [
      {"type": "message", "role": "user", "content": "Explain the CAP theorem simply."}
    ],
    "fusion": {
      "analysis_models": ["openai/gpt-5.2", "anthropic/claude-haiku-4-5", "zai/glm-5.2"],
      "synthesis_model": "anthropic/claude-sonnet-4-6"
    }
  }'

How it performs

We tested Fusion on DRACO, a deep research benchmark of 25 tasks stratified across 10 domains. Every Fusion configuration beat every solo model across the board, including the models each panel was built from.

Three findings stand out:

• Frontier quality at a quarter of the price. An all open-source panel with no frontier model anywhere in the stack scored statistically level with Claude Fable 5 at about 25 cents on the dollar

• Spend more, and the ceiling keeps rising. The premium panel beat Fable 5 outright by 8.5 points, the new best score on the benchmark

• The judge seat is cheap. Handing synthesis to Luna, the smallest model in the GPT-5.6 family, held frontier-judge quality. The panel is where quality is made and the judge is where cost is saved

Fusion also finished all 25 tasks in both runs, while the solo frontier model refused the same task twice, so uptime doesn't depend on any one provider's refusal policy either.

When your customers should use Fusion

Reach for Fusion when answer quality matters more than latency. It's no longer a premium trade-off: an all open-source panel matches frontier quality at a quarter of the price, so the extra model calls can cost less than one frontier call.

• Deep research and analysis: synthesize one credible answer from many sources with higher quality than any single model

• High-stakes work: get a second and third opinion on legal, medical, financial, and compliance calls before trusting one answer

• Combating hallucinations: resolve contradictions the panel surfaces toward the best-supported claim

• Cutting frontier spend: replace a solo frontier model with an open panel at ~25 cents on the dollar, without giving up quality

• Vendor fallbacks: keep answering even if a provider goes down, rate-limits you, or refuses

Fusion isn't the right tool for real-time, low-latency experiences, high-volume routine traffic, or streaming. Route those to a single fast model - latency, not cost, is now the reason to hold back.

Get started

Merge Fusion combines models into one answer that beats any of them alone, without frontier-level pricing. 

Schedule a demo to talk through your use case, or read our docs to make your first Fusion call in Merge Gateway.