Claude Science launched in public beta on 30 June 2026 and reached Windows on 10 September. Unlike almost every other "AI for research" product of the last two years, it is not a chat box with a literature search bolted on. It is a desktop application that installs Python and R environments on your machine, runs code, connects over SSH to your lab's cluster, and keeps a complete execution record of everything it did.
That distinction matters more than the feature list. This post covers what Claude Science actually is, the three or four things in it that are genuinely new, what it costs, what researchers found after a summer of using it, and the specific reasons it will not get a thesis written.
Writing the document, not running the experiment
ThesisAI drafts a full academic document from one prompt, with inline citations verified against papers retrieved from indexed databases, and exports to LaTeX, Word and BibTeX.
Inizia a scrivereWhat Claude Science Is
It is a downloadable app for macOS, Windows and Linux, included with a paid Claude plan rather than sold separately. Work is organised into Projects (a research topic) and Sessions (an individual line of analysis), and each project keeps its own memory, data sources, installed skills and artifacts.
Inside a session, you type a plain-language request. A coordinating agent takes it, decides what needs doing, and can spin up specialist sub-agents to do the parts. Those agents write and run code against data that stays on your own infrastructure. Anthropic's own description is careful on this point: your datasets do not leave your machine, only the context window is transmitted.
The headline capabilities break down roughly like this:
| Capability | What it means in practice |
|---|---|
| Reproducible artifacts | Every figure, table and result ships with the exact code, the environment that produced it, a plain-language description, and the full message history that led to it |
| Reviewer agent | A separate agent inspects outputs as they appear and flags incorrect citations, numbers it cannot trace, and figures that do not match their underlying code |
| 60+ database connectors | UniProt, PDB, Ensembl, Reactome, ClinVar, ChEMBL, GEO, OpenAlex, preprint servers, queried without learning each platform's interface |
| Compute that scales | Runs locally, or over SSH to a remote Linux box or a Slurm HPC cluster, or on Modal for GPU work, with a Compute tab showing every running kernel's memory and CPU |
| Persistent kernels | Python and R kernels keep their variables in memory across sessions, so you are not re-loading a 40 GB matrix every morning |
| Scientific renderers | 3D protein structures, genome browser tracks, chemical structures, sequence alignments, parquet and TSV as tables, PDFs, all rendered natively |
| Domain modules | Pre-configured for genomics, single-cell RNA-seq, proteomics, structural biology, cheminformatics, phylogenetics and CRISPR screens |
| Extensibility | MCP connectors for internal systems, plus reusable skills saved as SKILL.md files that persist across sessions and can be imported from private GitHub repos |
There is also an NVIDIA BioNeMo Agent Toolkit integration that puts Evo 2, Boltz-2 and OpenFold3 inside the environment, which is the sort of thing that tells you exactly which audience this was built for.
The Three Things That Are Genuinely New
Most of the above exists elsewhere in some form. Three things do not, and they are worth understanding properly.
1. The execution record is the artifact
In a normal AI-assisted analysis, you get a figure and a wall of code in a chat transcript, and reconstructing what produced what is your problem six months later. Claude Science inverts this: the figure is a bundle containing its code, its environment, and the conversation that produced it. You can reopen it much later and see precisely what was run.
This is the single most useful idea in the product, and it is a reproducibility idea rather than an AI idea. It would be valuable even if the model were mediocre.
2. A reviewer that reads the execution record
The reviewer agent is not a second opinion on your science. It is a consistency checker with access to the full run: it compares what the main agent said happened against what the record shows did happen. When a figure caption claims a filter was applied and the code shows it was not, that mismatch is exactly what it catches. Findings appear as cards under the message they refer to, and you can open the reviewer's reasoning.
Understand the boundary here, because it is the difference between useful and dangerous. The reviewer does not re-run your analysis and it does not evaluate whether the analysis was the right one. It verifies that the narrative matches the execution. A methodologically wrong analysis, faithfully described, passes review.
3. Compute planning as a first-class feature
The app drafts a plan before it reaches for a new resource, and waits for approval before running anything compute-intensive or billable. Sessions pause and ask before spending extra usage rather than quietly burning credits. For anyone who has watched an agent loop itself into a large bill, this is a meaningful design decision.
What It Costs
Claude Science itself carries no separate price. It requires a Claude Pro, Max, Team or Enterprise subscription, and it is included with those.
The more interesting offer is the academic one. Anthropic opened a Claude Team plan for scientists: 10,000 seats, standard seats free for a year, premium seats at $15 per month with 5x usage limits. Eligibility runs through the lab, not the individual - you have to be a principal investigator or equivalent at an academic or nonprofit research institution, and once verified you add your lab members to the plan. There is also an AI for Science credit programme for individual projects, open to any researcher rather than PIs only.
If you are a PhD student in a lab whose PI signs up, this is effectively free. If you are a master's student writing a thesis alone, you are on the normal consumer plans.
What Researchers Found After Using It
Early accounts from working scientists are more informative than the launch material, and they converge on a consistent picture: fast and transparent, but not a scientist.
Stephen Francis, a neuroscientist at UCSF, used it on glioma genetics datasets and found it suggested analysis steps he had not thought of: it would hit his own top three and add a couple more. Kelvin Lau at EPFL singled out the transparency, describing an app that shows you step by step what it is doing, which code it is using, and why. Jerome Lecoq at the Allen Institute built a multi-agent template that assembled literature reviews running past 100 pages with automated citation verification, far faster than a human would write them.
The criticisms are more specific than the usual complaints about hallucination.
- It does not doubt itself. Lecoq's point was that scientific methodology is not in the training objective: a model predicts the next token, and a scientist's job is to doubt it. The app is a fast executor of analyses, not a sceptic about them.
- Paywalls still stop it. It cannot read papers your institution does not have access to, which is the same wall every literature tool hits.
- It struggles to connect distant concepts. Reaching across fields to something that is not well represented in training data is where it thins out, and that is precisely where original contributions live.
- The biosecurity guardrails overshoot. Francis reported refusals on legitimate prompts about herpesviruses and brain tumours, and called the guardrails too stifling. Anthropic has acknowledged the tension in a roundabout way: choosing a model whose safeguards block most life sciences research now shows a warning first, with a one-click switch to an alternative model.
Nobody in these accounts described output they would submit unchecked. Anthropic does not claim otherwise either - the stated position is that results should be verified before use in research, publication or downstream decisions.
Where It Stops
Four limits determine whether Claude Science belongs in your workflow, and none of them is a bug.
The domain coverage is life sciences. Genomics, proteomics, structural biology, cheminformatics. If you are in economics, psychology, education, law or history, the pre-configured infrastructure that makes this app impressive simply is not there. You get a capable general model attached to a code runner.
There is no reference library. You can upload documents into a session, but there is no persistent library across projects, no Zotero or Mendeley import, no citation style engine. The citation checking it does is verification that a cited claim traces to something real in the run, not bibliography management.
There is no systematic review workflow. No PRISMA screening, no dual-reviewer blinding, no structured extraction tables with quote-level provenance. If your methods chapter has to survive a reviewer asking how records were screened, this app does not produce that audit trail. Our guide to systematic vs literature reviews covers what that actually requires.
It does not write a thesis. It drafts methods, results and figure captions alongside the analysis they describe, with Markdown and LaTeX preview that now resolves cross-references properly. That is genuinely useful and it is also the narrow slice of a thesis closest to the code. It does not plan chapters, build an argument across 80 pages, or take you from a research question to a structured document.
Claude Science vs ThesisAI
These tools sit at opposite ends of the same project, which is why the comparison is less a competition than a division of labour.
| Claude Science | ThesisAI | |
|---|---|---|
| Core job | Run and document the analysis | Draft and structure the document |
| Starting point | Your data and your compute | Your topic and your sources |
| Output | Reproducible artifacts, code, figures | A full document with inline citations, up to 80 pages |
| Citations | Verified against the execution record | Retrieved from indexed databases and verified against the source paper |
| Reference managers | Not supported | Zotero and Mendeley import |
| Best for | Wet-lab and computational researchers with data to analyse | Students and researchers with a document to produce |
| Setup | Desktop install, environments, connectors, compute config | Browser, one prompt |
If you are a computational biology PhD student, the honest answer is to use both: Claude Science for the analysis chapters, where reproducible artifacts are worth real effort, and a document-first tool for everything around them. We have a longer Claude Science vs ThesisAI comparison that goes feature by feature.
Who Should Install It
Install it if: you write code to analyse data as a routine part of your research; you work in genomics, proteomics, structural biology or cheminformatics; you have access to a cluster or GPU compute; you have ever failed to reproduce your own result from six months ago; your PI can get the lab onto the free academic plan.
Skip it if: your research is qualitative or theoretical; you are in the humanities or most social sciences; your bottleneck is writing and structuring rather than computing; you need PRISMA-grade evidence synthesis; you want something that works in a browser without installing an environment.
The second list is not a criticism. It is a description of an app that chose an audience and built properly for it instead of claiming to serve everyone.
FAQs About Claude Science
Is Claude Science free?
The app costs nothing extra but requires a paid Claude plan (Pro, Max, Team or Enterprise). Principal investigators at academic and nonprofit research institutions can apply for a free Claude Team plan for scientists, which covers their lab members for a year.
Does it work on Windows?
Yes, since version 0.1.47 on 10 September 2026. It launched on macOS and Linux only, so older articles will tell you otherwise.
Does my data get sent to Anthropic?
Analyses run on your own machine or your own remote compute, and Anthropic's position is that datasets stay there - only the context window is transmitted. If you handle restricted data, confirm this against your institution's data governance rules rather than a blog post, including any conditions in your ethics approval.
Can it replace my supervisor's feedback?
No, and the reviewer agent is a good illustration of why. It checks that your description matches your execution. It has no view on whether the design answers your research question, which is the thing a supervisor is for.
Is it a literature review tool?
Partly. It queries OpenAlex and preprint servers and can assemble long reviews with citation checking, and researchers have done exactly that. But it cannot read paywalled papers, has no reference library, and does not produce a screening audit trail, so it is not a substitute for a properly conducted review.
Can I cite Claude Science in my thesis?
You cite the papers, not the tool. Where the tool was used to produce analyses, that belongs in your methods section as a description of software used, alongside the versions of the packages it ran. Our guide to citing ChatGPT and AI tools covers the distinction between citing AI output and disclosing AI assistance.
Does it handle R, or is it Python only?
Both. Persistent Python and R kernels, with variables held in memory across sessions.
Claude Science is the first AI research product that looks like it was designed by people who have actually fought with a conda environment at 2am. The reproducibility model is the right one, the reviewer is a sensible and appropriately modest idea, and the compute handling is careful. It is also aimed squarely at people with data and a cluster, and a thesis is mostly not that. Use it for the analysis. Use something built for documents for the document.