DRACO, explained: how to benchmark deep research

Most language-model benchmarks test what a model knows. DRACO tests what a model can figure out — it measures deep research, not trivia recall. That makes it a useful yardstick for systems like Fusion API, where the whole point is reasoning and cross-checking, not memorized facts.

What DRACO measures

DRACO is a set of 100 tasks across 10 domains — academic, financial, legal, medical, technical, UX design, product comparison, and more. Each task goes beyond "what's the answer" to evaluate:

• Reasoning — can the model work through a multi-step problem?
• Tool use — does it search, retrieve, and use information well?
• Report quality — is the final write-up accurate, complete, and clearly expressed?

Every task carries close to 40 grading criteria, focused on factual accuracy, analytical completeness, information synthesis, citation reliability, and clarity.

Why you can't pad your way to a good score

Crucially, DRACO has negative scoring. Give a dangerous medical recommendation or cite obviously wrong information, and you lose points. That kills the usual trick of inflating a score by writing more — length without correctness actively hurts.

This is what makes DRACO a fair test for multi-model systems: the cross-review step in Fusion exists precisely to catch risky and unsupported claims before they reach the final answer. A benchmark that punishes those claims rewards that design.

How to read the numbers

A few caveats worth knowing if you compare DRACO results:

• Judge variance. Different judge models can shift absolute scores by 10–25 points, so don't compare raw numbers across different setups. Relative rankings stay stable, though.
• Coverage. If a model is blocked by content filters on some tasks, its score is computed over the tasks it completed — not always a like-for-like comparison.
• Search leakage. If a model happens to find the grading rubric online mid-task, that inflates its score. Good evaluations blacklist those pages and re-run.

The takeaway

On DRACO, a coordinated panel of efficient models lands within about a point of a single frontier model — at roughly half the cost. That's the empirical case for treating multi-model coordination as real infrastructure, not a gimmick.

See the scores on the home page, watch the process on the live demo, or join the waitlist.