Claude Science Is in Beta: Reproducibility Is the Feature That Matters

Anthropic has put a desktop research application into beta, and the most interesting thing on its page is not the headline. Claude Science bills itself as “your research partner for rigorous science,” and the sentence directly beneath that billing is about plumbing rather than intelligence: the app runs analyses, searches databases, and traces every step from data wrangling to publication. Downloads are offered for Mac on Apple Silicon, Mac on Intel, Windows and Linux, with a contact-sales button beside them for teams that want a conversation before an install. Nowhere on the page is the model named. Nowhere on the page is a price. What is named, in unusual detail, is provenance.

That ordering is worth sitting with. Most AI product pages lead with a capability and add trust language afterwards. This one leads with a record of how a result was made, and treats the analysis as the thing being sold: rich scientific artifacts that are fully reproducible, with every result traced to its code. If you run a lab, that is either the most useful sentence Anthropic has written this year or the one you should test hardest. The rest of this piece is about telling those two apart.

What the app actually is

Claude Science is an application in beta, not a service you reach through a browser tab. The download matrix — Apple Silicon, Intel Mac, Windows, Linux — says the assumption is that work happens where the data already sits, on a workstation or a lab machine. Anthropic’s own description of what the app does is narrow and specific: it runs analyses, searches databases, and traces every step from data wrangling to publication. The promise attached is not better answers, exactly. It is less time spent on the parts of research that are not research.

There is a second thing the page declines to do. It does not sell a subscription. The page carries a contact-sales button and no price, and the app is marked beta. The commercial detail is left to a conversation, which is a fact about the page rather than a complaint about the company — but it does mean that any number quoted as the price of Claude Science is quoting something that has not been published.

What the page does disclose, at some length, is the shape of the work. Every artifact ships with its history: the exact code, the environment, and the conversation that produced it. Anthropic’s phrasing is that figures, tables and notebooks “can be reproduced, edited, or defended months later.” That is a claim about record-keeping, and record-keeping is exactly the kind of claim a lab can check.

The artifact carries its own history

A figure in a submitted manuscript is a bitmap with an assertion attached. Six months later a reviewer asks why the error bars widen above a certain concentration, or a collaborator wants the same plot for a second cohort, and someone has to reconstruct a notebook last touched by a postdoc who has since left. The usual answer is archaeology: a shared drive, a chat thread, three versions of a script and a guess about which one produced the panels in the paper.

The claim here is that the artifact is not a picture but a record, and that the record travels with the picture. Code, environment and conversation, stored alongside the output, so that the person re-opening it in a year gets the same answer rather than a plausible one. If that holds, it changes the cost of a fairly ordinary research event: the revision request that arrives after everyone has moved on.

There is a second-order effect that matters more than convenience. If the code is attached, a figure can be edited rather than redrawn. Anthropic describes figure iteration in plain language: annotate a figure to request a change or ask a question, and the agent reads the code that produced it and edits directly. That is a different operation from regenerating an image. Redrawing a plot by hand introduces a fresh opportunity for an error nobody will catch, because the new plot looks like the old one. Editing the code that made it keeps the derivation intact — assuming, and this is the assumption the whole product rests on, that the code attached to a figure really is the code that produced it.

The reviewer that reads the code

Anthropic has built a background reviewer into the app. Its job is to flag three things: incorrect citations, numbers that cannot be traced to a source, and figures that do not match their underlying code. Each of those is a specific failure mode of computational biology rather than a generic hallucination, and the third connects directly to the provenance claim. A reviewer that can see both a figure and the code that made it can check one against the other. A reviewer that sees only the figure cannot.

What is not published is how well this works. There is no rate, no benchmark and no independent evaluation on the page, so none is offered here. The honest framing is that a reviewer which flags mismatches is a useful second pair of eyes and a poor substitute for a first pair. It will catch the figure whose axis labels drifted between versions. It will not tell you whether the experiment was designed correctly, and it cannot know that a sample sheet was mislabelled upstream if the mislabelling is consistent all the way through.

That distinction matters because the reviewer is the part of the product most likely to be oversold in a lab meeting. A green check from a background reviewer means a claim survived a consistency check, not that the claim is true. The provenance record is the more durable half of the pair: it does not depend on the reviewer being right, only on the record being complete.

Renderers, manuscripts and compute

Three smaller pieces of the app make the same argument from different directions. The first is native rendering. Proteins, alignments, genomic tracks, chemical structures and PDFs display in their native form with no extra installation. That sounds cosmetic until you remember how much of a lab’s toolchain is a chain of export steps — structure viewer to screenshot to slide to figure — each of which drops metadata and adds an opportunity for something to be silently resized.

The second is manuscript drafting. Results are written up alongside the analysis that produced them, with rendered Markdown and LaTeX previews. Keeping the prose in the same project as the computation is the cheapest available version of reproducibility: the draft and the analysis cannot drift apart if they never live apart.

The third is compute management. The app builds the environments each analysis needs and runs them on your laptop, your cluster, or GPUs. This is where a research tool either becomes usable or does not. A GPU job that has to be assembled by hand in a terminal is a job that only one person in the lab can run, and that person becomes a bottleneck the moment they take a holiday. Environment building is unglamorous, and it is the difference between a tool a lab adopts and a tool one graduate student uses.

Who gets it free, and who the page means by scientists

A banner sits above the product page advertising a Claude Team plan for scientists: scientists at academic and nonprofit research institutions around the world can now get Claude at no cost to start. That is Anthropic’s wording on Anthropic’s own page, and its boundaries matter as much as its content.

The offer is framed around academic and nonprofit research institutions. It is not written for industry R&D and not for students as a category; extending it to either would be reading it beyond what the sentence says. The phrase “to start” is doing work too. The page describes a no-cost start, not a permanent entitlement, and it does not say what comes next. A lab planning a budget around this should read it as a door-opener rather than a line item, and should not assume the app itself has a published price, because it does not.

The questions the page leaves open

Five things a careful reader will look for and not find. The model: the page does not say which Claude model or models Claude Science runs on, so any article that names one is guessing. The price: no subscription is sold on the page. System requirements: four platforms are listed and no hardware or OS versions are given. Whether artifacts leave the machine: the provenance story is about what is stored with a result, not about where that result is sent, and those are different questions. And an exit date from beta: none is published.

The fourth of those is the one to resolve before the first upload. A reproducibility record is valuable precisely because it is detailed — code, environment, conversation — and detailed records of unpublished research are precisely what a data-governance committee will ask about. The page does not answer where the record lives, and that silence is a legitimate reason to have the conversation with Anthropic before putting clinical data through the app rather than after.

The other announcement that week

On 22 October 2026 Anthropic published an open call for guest investigators in the life sciences, inviting the life-sciences community to apply to work with Anthropic on special projects. That is the entire published content of the announcement. It is not a programme description, not a grant, and not a statement about Claude Science’s roadmap; treating it as any of those would be reading in more than is there. It is, though, a signal about where the company is putting attention. Anthropic is recruiting practising scientists, not customers.

What this means if you are building rather than researching

One boundary first. Claude Science is an Anthropic application. We do not host it, we do not sell it, and nothing here should be read as an availability claim for the app. What we host is models, through one OpenAI-compatible endpoint with a single key, at provider list price with no markup added on top. Claude Opus 5.5 sits on that catalogue at $4.00 per million input tokens and $20.00 per million output tokens, with a 1M-token context and cache reads at $0.20 per million.

That is relevant to a narrower slice of this story than it first appears. Because Anthropic has not published which model Claude Science runs on, there is no honest way to say the app runs on a model we happen to carry. What can be said is that the parts of a research pipeline which live outside the app — a batch job that annotates samples overnight, a script that calls a model from inside an existing notebook, a step that has to keep running through a provider outage — are ordinary model calls, and those are what a router is for. Automatic failover across providers matters more in a job that runs for six hours than in a chat window that a person is watching. The routing DSL, which composes several models into a single call, matters when one step in a pipeline wants a small cheap model and the next wants a careful one.

The reproducibility discipline the app is selling does not have to stop at the app’s boundary. If a result depends on a model call you made yourself, the same rule applies: record the model, record the prompt, record the date. Endpoints outlive the versions behind their aliases, and a figure that depends on a call nobody documented is exactly the figure that becomes archaeology.

The test worth running first

Every product page in this category makes a version of the same promise: the work will be faster and the record will be cleaner. The first half is easy to demo and hard to verify. The second half is verifiable, and Claude Science has made it unusually specific — code, environment and conversation, attached to the artifact, defensible months later.

So the test is not a demo. Take a figure your lab has already published, one that took real work to produce, and ask the app to reproduce it from the record. If it comes back with the same numbers and a readable account of how they were reached, the provenance claim is real for your kind of work. If it comes back with a plausible-looking plot and no trail, you have learned something more useful than any benchmark: the tool is a renderer with good manners, and your lab still needs its own archive.

Either answer is worth having. A beta is a hypothesis about how research will be done, and the helpful thing about this particular beta is that its central claim is falsifiable by anyone with a published figure and an afternoon.