Executive Summary

A general chatbot summarizes papers and suggests code. But actual research is far more fragmented than that. A researcher hunts for literature in PubMed, analyzes in Jupyter and R, logs into a cluster terminal, and hops several times a day between dozens of life-science databases, incompatible file formats, and visualization tools. Every jump between tools breaks context, and those breaks pile up until they erode productivity and reproducibility together.

The scale of that fragmentation shows up in the numbers too. One survey found that a research organization manages more than 100 data sources on average, and 30% of them handle over 1,000. There's also a long-cited rule of thumb that a large share of research time — estimates range from 45% to 80%, with wide variation by source — is spent just getting data ready to analyze. The problem isn't the model's intelligence; it's that the tools are scattered and no thread of trust runs between them.

There's a misconception to clear up first. Claude Science is not a new AI model. It isn't a model trained harder for biology, either. Anthropic is explicit that it is a beta app running on the same existing Claude models anyone uses today (including Claude Opus 4.8), with no gating or special access. What's new isn't the model's brain but the tools that brain now holds in its hands — research skills, 60-plus database connections, a code-execution environment, HPC integration, and reproducibility records.

The structure is easiest to grasp in three layers. As in the diagram below, at the bottom sits the unchanged existing Claude model; above it a research-tooling layer (60+ skills, connectors, and databases); and above that an execution-environment layer (local, SSH, HPC, Modal, plus data locality).

The orchestration is distinctive too. A main AI splits work like a project manager (a PI) and delegates to sub-assistants, while a separate reviewer (fact-checker) agent re-checks the citations, calculations, and figures behind the results. Ask a question in plain language and Claude finds the literature, queries databases, runs analysis code, generates figures, and carries the thread all the way to a manuscript draft — within the same flow. The product's stated aim is not "automating discovery" but "integrating scattered tools."

Claude Science's biggest differentiator is provenance. Every figure, table, and notebook it produces carries along the exact code, execution environment, plain-language description, and full conversation history that created it. So months later you can still trace "which data and which code made this figure." Why that matters becomes clear when you look at an old wound in science.

3.1The Reproducibility Crisis as Backdrop

When Nature surveyed some 1,500 scientists in 2016, more than 70% said they had failed to reproduce another lab's experiment. The failure rate varies by field, but it's low nowhere. Below are those failure rates by discipline.

Share reporting they had "failed to reproduce another lab's experiment," by field (Source: Baker, Nature 533, 2016, n≈1,576)

The cost is steep as well. Irreproducibility in U.S. preclinical research alone is estimated to tie up about $28 billion a year (Freedman 2015), and when the drugmaker Amgen re-ran 53 landmark cancer studies, the core result reproduced in just 11% (6 studies) (2012). The 2021 cancer-biology reproducibility project reported that 59% of the targeted experiments failed to replicate, and even the effects that did replicate were a median of 85% smaller than the originals. A result can exist, but if the process that made it is gone, science cannot check itself.

3.2The Irony of AI as Both Problem and Cure

A new source of contamination has now been added: fabricated citations generated by LLMs. The rate of this phenomenon — inventing plausible-sounding references to papers that don't exist — spans 14% to 95% depending on model and domain, and even well-known early studies found GPT-3.5 fabricating 55% of its citations and GPT-4 fabricating 18% (Walters & Wilder 2023). Bolt on retrieval augmentation (RAG) and 3–13% still slip through. A large 2026 Lancet audit found the fabricated-citation rate in biomedical papers had climbed more than 12-fold from 2023, so that by early 2026 one paper in 277 cited a reference that doesn't exist.

Claude Science's reviewer agent aims squarely at this contamination. It's designed to catch and flag wrong citations, numbers with no traceable source, and figures that don't match their code — before publication. In the very place where AI pollutes the literature, another AI tries to filter that pollution out. The verification flow looks like this.

If provenance is the axis of trust, the remaining values are the axis of actually letting the researcher get the work done: HPC in plain language, execution where the data already lives, coverage pre-tuned for the life sciences, and an unbroken line from analysis to manuscript.

4.1HPC in Plain Language

Claude Science can run not only in local environments but through Linux servers, HPC login nodes, SSH-based clusters, and Modal accounts. By Anthropic's account, it writes batch scripts and, over SSH, submits and manages jobs on your own machine or an HPC cluster, scaling from a single GPU to hundreds. There's a stated safeguard as well: you can review, approve, or withdraw a plan before a job is submitted. For a non-computational researcher who isn't fluent in Slurm, SSH, or conda, this is an attempt to "lower the barrier to HPC through a natural-language interface."

That this barrier is no exaggeration shows up in the numbers. In a survey of ML researchers, 62% could access fewer than 8 GPUs, and 57.4% said they had been unable to run an experiment at all for lack of compute (among ML researchers, arXiv 2306.16900). There's no equivalent figure carved out for life scientists specifically yet, but the reality that simply putting one job on a cluster is a threshold isn't much different. If a natural-language interface lowers that threshold, the payoff comes not because the model got smarter but because access got wider.

4.2It Runs Where the Data Lives

From a governance standpoint, the most decisive design choice is data locality. Anthropic explains that because Claude Science runs on a lab's own infrastructure — laptops, Linux boxes, HPC login nodes — large or sensitive datasets don't need to leave the systems they already sit in. Only the context each analysis step needs is sent to Claude. There's a caveat attached, though: whatever ends up in prompts and model responses is handled under Anthropic's standard retention policy. If you work with regulated data such as HIPAA or clinical records, that caveat is one to confirm carefully.

4.3Pre-Tuned for the Life Sciences

Claude Science ships more than 60 curated skills and connectors tuned to the major life-science fields — genomics, single-cell analysis, proteomics, structural biology, cheminformatics. It can query more than 60 scientific databases, with resources like UniProt, PDB, Ensembl, Reactome, ClinVar, ChEMBL, and GEO offered as examples. Connect it to the NVIDIA BioNeMo Agent Toolkit and it reaches past tools to the latest research models themselves. Evo 2 is a genome foundation model trained on roughly 9.3 trillion base pairs spanning some 128,000 species; OpenFold3 is an open implementation of the protein-structure predictor AlphaFold3; and Boltz-2 predicts molecular binding affinity. Flip that around and the coverage is heavily biased toward the life sciences — generalizing to physics or the social sciences still calls for caution.

4.4Reusing Existing Pipelines, and Figure-to-Manuscript Continuity

Claude Science doesn't ask you to throw away the tools you have. You can bring your existing Python, R, and shell workflows as-is, and connect validated ELNs and internal systems as connectors or skills. When the analysis is done, it flows straight into figures and a manuscript. It natively renders 3D protein structures, genome-browser tracks, and chemical structures, and when you make a plain-language request like "switch the axis to log scale," it edits the code that made the figure directly. From data query to manuscript, the flow never breaks.

The race in science AI now splits two ways. One camp aims at powerful, dedicated models that automate discovery itself. OpenAI's GPT-Rosalind is a dedicated model specialized for biological reasoning, placed behind enterprise gating; Google's Gemini for Science takes the form of a desktop workbench, competing directly with Claude Science. FutureHouse pursues an autonomous-scientist agent. The table below plots each product's coordinates by its approach philosophy. Read it as a difference in design philosophy, not a ranking.

On this map, Claude Science's coordinates are distinct. It weighs verifiability over benchmark scores. Rather than making discoveries faster, it leans toward making the discoveries you've made reproducible and auditable. And this is exactly where the Pebblous lens locks in.

The thesis Pebblous has long argued is "data over model." Claude Science's provenance — "which data and which code made this figure" — is isomorphic to the problem Pebblous DataClinic tackles when it diagnoses data quality and traces lineage. The "wrong citations, numbers with no traceable source, code–figure mismatches" the reviewer agent catches are textbook data-quality defects of consistency, lineage, and verifiability. No matter how large the model, you can't trust the output without trust in the input and intermediate data — and the fact that citation fabrication persists even in GPT-4-class models shows exactly that.

The hotter the "AI that automates discovery" race gets, the more the bottleneck moves from the speed of discovery to its verifiability and reproducibility. For life-science, pharma, and materials R&D customers, the demand that "AI runs the pipeline for us, but the data never leaves our infrastructure and every result carries its lineage" is, at heart, a demand for data governance and quality. Claude Science is a flagship case of data-lineage management becoming a first-class product feature — outside proof of why the "data-trust infrastructure" Pebblous has been describing is needed.

Primary product source

• 1.Anthropic (2026, June 30). Claude Science, an AI workbench for scientists. anthropic.com

Reproducibility crisis

• 2.Baker, M. (2016). 1,500 scientists lift the lid on reproducibility. Nature, 533, 452–454. doi.org/10.1038/533452a
• 3.Freedman, L.P., Cockburn, I.M., & Simcoe, T.S. (2015). The Economics of Reproducibility in Preclinical Research. PLOS Biology, 13(6):e1002165. doi.org/10.1371/journal.pbio.1002165
• 4.Begley, C.G., & Ellis, L.M. (2012). Raise standards for preclinical cancer research. Nature, 483, 531–533. doi.org/10.1038/483531a
• 5.Errington, T.M. et al. (2021). Investigating the replicability of preclinical cancer biology. eLife, 10:e71601. doi.org/10.7554/eLife.71601

LLM citation hallucination

• 6.Walters, W.H., & Wilder, E.I. (2023). Fabrication and errors in the bibliographic citations generated by ChatGPT. Scientific Reports, 13:14045. doi.org/10.1038/s41598-023-41032-5
• 7.Chelli, M. et al. (2024). Hallucination Rates and Reference Accuracy of ChatGPT and Bard for Systematic Reviews. Journal of Medical Internet Research, 26:e53164. doi.org/10.2196/53164
• 8.Topaz, M. et al. (2026). Large-scale audit of fabricated references in the biomedical literature. The Lancet, 407:1779–1781.

Compute & data gaps · life-science foundation models

• 9.Gao, S. et al. (2025). AI for Scientific Discovery is a Social Problem. arXiv:2509.06580. arxiv.org/abs/2509.06580
• 10.Stephens, Z.D. et al. (2015). Big Data: Astronomical or Genomical? PLOS Biology, 13(7):e1002195. doi.org/10.1371/journal.pbio.1002195
• 11.Brixi, G. et al. (2025). Evo 2: Genome modeling and design across the tree of life. (preprint) biorxiv.org
• 12.Wohlwend, J. et al. (2025). Boltz-2: Accurate and Efficient Binding Affinity Prediction. bioRxiv 2025.06.14.659707. doi.org/10.1101/2025.06.14.659707
• 13.arXiv:2306.16900 (2023). Barriers to compute access in ML research. arxiv.org/abs/2306.16900 (62% of ML researchers had fewer than 8 GPUs; 57.4% couldn't run experiments for lack of compute.)

Industry & market

• 14.TechCrunch (2026, June 30). Anthropic bets on workflow, not a new model, to win over scientists.
• 15.Grand View Research (2026). Artificial Intelligence in Drug Discovery Market Report.

Market-size estimates vary by up to 10× across firms because of definitional differences, so this piece cites ranges and sources rather than asserting a single figure. Claude Science efficiency claims (e.g., time saved on a specific analysis) come from Anthropic announcements and users' self-reported figures, and are not independently verified.