Comparing AI research tools on 95 life-science questions

We gave Arca, Elicit, Consensus and SciSpace the same 95 life-science research questions. SciSpace had the highest average rubric score at 84.81, followed by Elicit at 82.35, Arca at 79.45 and Consensus at 56.13.

These scores describe a mixed-version snapshot collected from October 29 through November 3, 2026. Arca used older local builds for 70 answers and production for 25. Product modes varied too. The overall score is not a measurement of current production Arca. We ran the evaluation ourselves and used AI judges; the scientific answer keys have not had a full expert review.

Results

Every score in the chart covers the same 95 questions. Each question has equal weight in the average. Scores measure rubric points out of 100, not the percentage of questions answered correctly.

Undermind scored 71.63 on a 20-question subset using its free plan. It has no score for the full dataset, so we leave it out of the chart rather than compare different questions.

The overall average hides differences between kinds of work. On survey-based questions, Consensus scored 82.54, SciSpace 80.93 and Elicit 80.68. Their small gaps do not establish a winner or prove equivalence. Consensus scored much lower on questions about recent regulatory actions and company claims, which can require sources beyond journal articles.

SciSpace also beat Arca on the questions we ran in production, scoring 91.40 against Arca's 83.74. That paired gap was 7.66 points, with a standard error of 1.80 across 25 questions. We did not rerun Arca to replace that loss.

The dataset

The questions ask for research work a life-science team might need during a project. Some require a synthesis of papers. Others ask whether a trial registry agrees with a publication, whether a paper was retracted, or what a regulator actually decided about a drug or device.

We wrote 45 questions from recent events and primary sources. The other 50 come from ResearchQA, whose answer criteria draw on survey articles. We fixed the prompts and reference answers before the product runs. No product received the grading criteria or answer key.

This is a different task from the regulatory deliverables in LifeSci Diamond. That benchmark also compares plain model APIs. The scores are not interchangeable.

Product settings

A vendor can change its models without changing the interface label. These results cover the selected modes and dates, not every feature each product offers.

We kept the finished answers, generated reports and citation references. Downloaded source files, search snippets and visible reasoning did not count as answers. Some bibliographies appeared outside the final chat message. We recovered those references from the saved thread before grading, without asking the question again. A report that repeated the final answer counted once.

Grading and uncertainty

The jury used openai/gpt-6-astra, anthropic/claude-opus-5 and google/gemini-3.1-pro-preview. Each model graded each rubric line. We took the median verdict, applied that line's weight and subtracted any triggered penalties. Question scores range from zero to 100.

The reference answers still need scientific review. A survey article can omit a valid answer found elsewhere, and judges can miss a citation error or prefer a particular writing style. We chose the questions, built the grading process and operated the accounts. This is not an independent evaluation of Arca.

Two ResearchQA keys already had flagged concerns. We kept both in the main result rather than change the keys after seeing product answers. Removing them in a separate check leaves 93 questions and gives SciSpace 85.50, Elicit 83.03, Arca 80.20 and Consensus 55.73. That check does not validate the remaining keys.

Most results use one answer per product and question. In separate repeats, Arca and Elicit answers moved about 5 points per question between runs on average. Small score gaps need more evidence than one run. The standard error quoted above measures variation across the selected questions, not uncertainty in the keys or AI judges. We retained refusals and product failures, including a Consensus failure.

We have not published the full keys, product captures or scoring harness for this comparison. Readers cannot yet reproduce these scores from public files. The public LifeSci Diamond harness belongs to the separate benchmark.

Completion times include polling and capture work, and some threads ran concurrently. We do not use them to rank product speed.

What we spent

All amounts are US dollars. SciSpace's total includes $20 for Premium and a prorated $58.48 Advanced upgrade. Elicit's extra usage covered 39 questions, about $1.29 each before the subscription. Its subscription also funded development and repeat runs.

These are experiment expenses, not advertised per-query prices. They exclude Arca operating costs and judge-model costs, so they do not establish which product is cheapest. We cancelled renewals and turned off Elicit extra usage after the runs.