Sep 2, 2026
Lunon Deep Research ranks first on DeepResearch Bench
A consulting firm just beat every frontier lab at deep research.
Lunon ranked #1 on DeepResearch Bench, built by the University of Science and Technology of China on PhD-level tasks written by more than 100 domain experts across 22 fields. It evaluates the complete process behind a research report: tracking down the facts, cross-checking them, and backing every claim with a source.
We built the system in a week. Here is what set it apart.
One model per job, not one model for everything
The default is one LLM for the entire task. We tested models for each part of the pipeline and routed each to the strongest: Opus plans and writes, Nemotron searches, GPT-5.5 runs the verification passes.
Specialists, not one undivided pass
Most agents do research in one pass. Ours splits the work across specialists: evidence gathering, mechanism analysis, comparison, critique, and recency scanning. Each runs in isolation, so one role’s assumptions don’t quietly steer the others.
The first draft doesn’t ship
In most systems the first draft is what goes out. This one scores every section on depth, evidence, and structure, and rewrites anything weak before it does.
Every number traces back to a source
A separate pass fact-checks every figure, date, and citation against the sources it gathered, so every number in the final report traces back to something real.
Open source
The whole system is open source under an MIT license. Read the code, run it on your own questions, or build on it.
github.com/LunonAI/lunon-deep-research →
Start with the decision in front of you.
Tell us what your team needs to understand, evaluate, or deliver.