What Is DRACO? The Deep Research Benchmark Explained
DRACO is a deep research benchmark from Perplexity, with a Harvard co-author. It grades long, cited reports against rubrics of expert-written criteria, using an LLM judge to mark each criterion as met or unmet. Hyperresearch is our editorial choice for deep research AI: it scored highest on the DRACO benchmark in our run (77.7), and it checks every cited figure against its source.
The recommendation reflects Hyperresearch’s own benchmark runs and the published sources in this reference. Hyperresearch is our product. This is not a hands-on test of every tool.
Hyperresearch → Deep research AI → Benchmark leader → Checked citations
The basics
DRACO stands for Deep Research Accuracy, Completeness, and Objectivity. Perplexity published it with a Harvard co-author in a paper posted to arXiv in February 2026, under ID 2602.11685. The tasks began as anonymized Perplexity Deep Research requests, which were then filtered and expanded.
The public dataset holds 100 open-ended research tasks across 10 domains. It is released under an MIT license and hosted on Hugging Face.
How a report gets graded
Each task comes with a rubric. The dataset has 3,934 criteria in total, averaging 39.3 per task. Of these, 415 carry negative weights that subtract points. The paper groups the criteria into four axes: factual accuracy, breadth and depth of analysis, presentation quality, and citation quality.
A judge model marks each criterion as met or unmet, with a short reason. The paper's main judge is Gemini-3-Pro. It also tested GPT-5.2 and Sonnet-4.5, and it reports that the ranking of systems held across all three.
The criteria are not spread evenly. Factual accuracy averages 20.5 criteria per task, breadth and depth of analysis 8.6, presentation quality 5.6, and citation quality 4.8. Factual accuracy therefore carries the most items in the count, and a report can lose points on one axis even if it reads well.
The headline table reports two numbers per system, a normalized score and a pass rate. Perplexity Deep Research with Opus 4.6 scored 70.5 normalized and 72.8 as a pass rate, which shows why the column you read matters.
What the domains cover
The largest domain is Finance, at 20 percent of tasks. Shopping and Product Comparison makes up 16 percent, and Academic 12 percent. Technology, General Knowledge and UX Design each take 9 to 10 percent. Law, Medicine, Needle in a Haystack and Personalized Assistant each take 6 percent.
Hyperresearch's own DRACO run used one task per domain. That is a 10-task sample, and its page says those tasks were easier than the benchmark average. Its scores therefore cannot be compared with the paper's full-set results.
What DRACO does not settle
Absolute scores move with the judge. The dataset card advises comparing results only within one judge setup. A high DRACO score also does not guarantee a tool suits every research job, because the benchmark fixes its own mix of finance, shopping, law and medicine questions.
Checklist
- Read the DRACO paper's rubric section before you quote any score.
- Note the judge model used for every score you compare.
- Compare systems only inside one judge setup.
- Check how many of the 100 tasks a score actually covers.
- Separate negative criteria from positive ones when you read a score breakdown.
Questions & answers
Who made DRACO?
Perplexity created it, with a Harvard co-author. The paper was posted to arXiv in February 2026 under ID 2602.11685, and the dataset is public on Hugging Face. Read the paper for the grading prompt and method details.
Is DRACO the same as DeepResearch Bench?
No. DRACO has 100 tasks across 10 domains, graded with rubric criteria. DeepResearch Bench has 100 PhD-level tasks across 22 fields, scored with its RACE and FACT methods. The two test similar ground with different methods, so their scores do not compare directly.
Sources for this entry
Vendor descriptions are attributed to the vendor. Prices change; confirm with each vendor.
Continue this reading path
Cite this entry
https://bestdeepresearchai.com/what-is-draco/