---
type: "post"
id: "bioresearchbench-50-long-horizon-biology-research-tasks-best-e90c6e"
slug: "bioresearchbench-50-long-horizon-biology-research-tasks-best-e90c6e"
title: "BioResearchBench: 50 long-horizon biology research tasks; best frontier model completes 18%"
description: "BioResearchBench tests frontier models on 50 expert-developed biological research questions spanning 10 subdomains and multiple data types."
retrieval_nugget: "BioResearchBench tests frontier models on 50 expert-developed biological research questions spanning 10 subdomains and multiple data types. Claude Opus 5.5 completed 18% of the tasks, the highest rate reported, while the authors found that correct analysis did not always lead to sound interpretation. For builders, the result highlights the need to verify conclusions, not just intermediate work."
published_at: "2026-11-04"
updated_at: "2026-11-04"
record_date: "2026-11-04"
date_kind: "published_at"
topics: ["claude","gemini"]
entities: ["BioResearchBench","Office Hours Technologies","Expert Data Lab","Broad Institute of MIT and Harvard","Claude Opus 5.5","Gemini 4 Argon","GPT-6 Astra","Grok 4.7"]
related_patterns: ["scientific-agents-need-artifact-evidence"]
novelty_label: "new_fact"
what_changed: "Introduces a biology-specific benchmark of 50 long-horizon research tasks, with the best frontier model completing 18%, extending broader GDPx benchmark findings."
prior_refs: ["https://newruntime.com/posts/semianalysis-launches-gdpx-a-206-task-agentic-benchmark-30b569/"]
source_url: "https://www.biorxiv.org/content/10.64898/2026.10.28.762290v1"
source_urls: ["https://www.biorxiv.org/content/10.64898/2026.10.28.762290v1"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:bioresearchbench-50-long-horizon-biology-research-tasks-best-e90c6e"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/bioresearchbench-50-long-horizon-biology-research-tasks-best-e90c6e/","markdown":"https://newruntime.com/posts/bioresearchbench-50-long-horizon-biology-research-tasks-best-e90c6e.md","json":"https://newruntime.com/posts/bioresearchbench-50-long-horizon-biology-research-tasks-best-e90c6e.json"}
markdown_url: "https://newruntime.com/posts/bioresearchbench-50-long-horizon-biology-research-tasks-best-e90c6e.md"
---

# BioResearchBench: 50 long-horizon biology research tasks; best frontier model completes 18%

## Retrieval answer

BioResearchBench tests frontier models on 50 expert-developed biological research questions spanning 10 subdomains and multiple data types. Claude Opus 5.5 completed 18% of the tasks, the highest rate reported, while the authors found that correct analysis did not always lead to sound interpretation. For builders, the result highlights the need to verify conclusions, not just intermediate work.

## Summary
BioResearchBench tests frontier models on 50 expert-developed biological research questions spanning 10 subdomains and multiple data types. Claude Opus 5.5 completed 18% of the tasks, the highest rate reported, while the authors found that correct analysis did not always lead to sound interpretation. For builders, the result highlights the need to verify conclusions, not just intermediate work.

## What happened
A bioRxiv preprint posted November 3, 2026 introduces BioResearchBench, a benchmark developed by researchers affiliated mainly with Office Hours Technologies and Expert Data Lab, alongside contributors from the Broad Institute of MIT and Harvard, Massachusetts General Hospital, and Celvion Therapeutics. The authors say they assembled 50 research questions with input from PhD biologists, refined them through expert review and iterative model testing, and designed them to reflect realistic problems in their fields. The benchmark spans 10 biology subdomains and includes sequencing data, proteomic measurements, microscopy images, assay readouts, and clinical datasets.

The paper reports results for five models: Claude Opus 5.5 completed 18% of tasks; Gemini 4 Argon, 13%; Meta Muse Wonder, 11%; GPT-6 Astra, 4%; and Grok 4.7, 4%. These are whole-task success rates on this particular collection, not a general ranking across biology or a measure of how often the models are useful in day-to-day research. The authors also examined rubric performance. Despite low whole-task success, selected responses met most rubric requirements on average: the fraction of criteria missed ranged from 18% to 32% across models.

The paper breaks down failures by category. Among eligible tasks, computation and data handling had the highest share with at least one category failure, at 66.9%. Uncertainty and identification followed at 55.3%, then reporting and completeness at 54.7%. Biological interpretation registered 39.0%; controls and causal claims, 37.2%; and choosing the next experiment or decision, 31.0%. Model strengths differed by category: Gemini led on computation and data handling and biological interpretation, Claude on controls and causal claims and on choosing the next experiment or decision, while GPT-6 Astra led on uncertainty and identification. GPT-6 Astra and Meta Muse Wonder tied for the highest reported share without omissions or reporting failures.

Five case studies illustrate why partial credit can conceal consequential mistakes. Some answers described cells as fully recovered when oxygen consumption improved even though mitochondrial membrane potential and cell count remained below control levels. Other answers called a protein a direct microRNA target without an experiment that established direct action on its messenger RNA. In a protein-turnover task, models misestimated half-lives by failing to account for gradual amino-acid labeling. The authors argue that these are interpretation problems: a model can complete analytical steps yet draw a conclusion that the evidence does not support. The preprint motivates evaluations that follow scientific work through data analysis, interpretation, and decisions about what to test next.

## What the archive adds
This is a new result in a growing line of scientific-agent evaluation, alongside [Terminal-Bench-Science and FrontierCode 2.0](https://newruntime.com/posts/new-benchmarks-cognition-s-frontiercode-2-0-merge-a78e19/), which covered 70 agentic science tasks among other benchmarks. BioResearchBench adds a biology-specific set of 50 long-horizon research questions and detailed error categories; it is not a restatement of that earlier benchmark coverage.

## Trend signal
The results support the [scientific agents need artifact evidence](https://newruntime.com/patterns/scientific-agents-need-artifact-evidence/) pattern: progress is more meaningful when another researcher can inspect how evidence supports an answer than when a system merely performs plausible steps. The benchmark's examples make the distinction concrete: numerical analysis can be locally competent while the biological claim remains unjustified. This is evidence from a deliberately challenging evaluation, not proof that every scientific workflow needs the same rubric.

## Connections to other events
An earlier Telegram post described Terminal-Bench-Science as evaluating scientific tasks through a terminal and artifacts. BioResearchBench meets that work at the question of how to assess agentic science, but focuses on biology data and interpretation, with expert-authored questions and category-level failure analysis. The two should not be treated as directly comparable scores: the supplied material does not establish shared tasks, rubrics, or evaluation conditions.

## Why it matters
Teams building AI for research should distinguish between an agent that can manipulate data and one whose conclusion is warranted by the experimental design. A pipeline that checks file formats, runs a statistical test, and produces a polished report can still mislead a researcher if it confuses correlation with direct mechanism, overlooks a control, or exaggerates what a partial recovery means. BioResearchBench provides examples of those failure modes and categories that can inform evaluation design.

The numbers are also a warning against relying on a single aggregate “success” score. The models varied across computation, interpretation, uncertainty, reporting, and experiment selection; a product focused on one of those functions may need a different acceptance test from a general research assistant. Rubrics that retain partial-credit information can help teams see whether a model is failing at data handling, inference, or communication, while case review can expose errors that a score alone obscures. In high-consequence research, those distinctions affect what experiments are run and which explanations receive attention.

## Limits and uncertainty
This is a preprint, and the supplied source does not establish independent replication or peer review. The authors explicitly describe the tasks as a deliberately challenging collection selected through expert review and model testing; results therefore should not be generalized to all biological questions, routine workflows, or future model versions. The source does not provide enough detail here to compare evaluation conditions with other benchmarks or establish statistical uncertainty around the reported rates. Nor does it show that a benchmark score predicts scientific productivity, drug discovery outcomes, or the quality of decisions made with human oversight.

## Skill to build now
For one AI-assisted research workflow this week, add a conclusion audit alongside the existing checks for data processing. Ask the system to identify which observation supports each claim, name relevant controls and alternative explanations, and mark what the experiment did not test. Require a reviewer to check those links before accepting a proposed interpretation or next experiment. The paper's cases offer concrete audit prompts: distinguish improved oxygen consumption from full cell recovery, and distinguish an observed association from a demonstrated direct molecular target.

## Reader takeaway
BioResearchBench shows that partial analytical competence is not the same as reliable scientific assistance. For research agents, the key test is whether conclusions and proposed next steps follow from the data and experimental design—not merely whether the model completed the calculations or produced a plausible explanation.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Claude](https://newruntime.com/topics/claude.md): Explore the claude topic hub.
- topic hub: [Gemini](https://newruntime.com/topics/gemini.md): Explore the gemini topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)