ResearchQA leaderboard for AI research tools in life sciences
This leaderboard ranks AI research tools and frontier models on the same 50 life-science questions from ResearchQA, a public benchmark from the University of Pennsylvania. We rerun it when a new model or product version ships. Last updated November 6, 2026.
Arca leads with 84.50 out of 100. Gemini 3.1 Pro and GPT-6 Astra with web search come next. All three score above the dedicated research tools Consensus, SciSpace and Elicit.
| Rank | System | Type | Score |
|---|---|---|---|
| 1 | Arca | Life-science research agent | 84.50 |
| 2 | Gemini 3.1 Pro | Model with web search | 83.53 |
| 3 | GPT-6 Astra | Model with OpenAI web search | 83.14 |
| 4 | Consensus | Research tool | 82.54 |
| 5 | SciSpace | Research tool | 80.93 |
| 6 | Elicit | Research tool | 80.68 |
What we learned
Frontier models have caught up with the research tools. A plain call to Gemini or GPT-6 with web search now writes a literature review that scores higher than Consensus, SciSpace or Elicit.
The gaps at the top are small. Single questions move about 5 points between repeat runs, so the first three places are close. Arca's lead over SciSpace and Elicit is larger and holds across the questions. For a head-to-head with an example question, see Elicit vs Consensus vs SciSpace vs Arca on ResearchQA.
Paper answers are only part of life-science work. ResearchQA asks literature questions. Regulatory, clinical and quality teams also need labels, recalls, registry records and their own documents. Arca reads more than 250 public sources and your files in Veeva Vault and SharePoint, which is what the LifeSci Diamond benchmark tests.
How we test
The 50 questions are ResearchQA's life-science and pharmacy questions, covering biochemistry, oncology, microbiology, pharmacology, toxicology and more. Each asks for a short literature review and has a rubric of 6 to 8 points drawn from published survey articles. No system saw the rubric.
Each system answered each question once in a fresh session. We used Elicit's Pro research reports, SciSpace's Pro agent, Consensus Pro and Arca in production on its Max model tier. The models ran with web search and a short instruction to answer in markdown and cite sources.
Three AI judges graded every rubric point. We used openai/gpt-6-astra, anthropic/claude-opus-5 and google/gemini-3.1-pro-preview and took the median. Two of the judges come from the same companies as two of the models we test, which may favor those models slightly. We built and run this leaderboard ourselves.
Changelog
- November 6, 2026. First version. Six systems on 50 questions. Research tool answers collected October 29 to November 5, model answers on November 6.
More comparisons
