Anthropic Launches Claude Science, Claims 10x Speedup in Scientific Research

Two years' worth of work done in just a few weeks.

Recently, neuroscientist Jérôme Lecoq and his team at the Allen Institute reduced the writing time for a lengthy review article from nearly two years to just a few weeks.

Jérôme Lecoq has accumulated around ten review papers, many exceeding 100 pages, with every citation verified sentence by sentence by an agent.

The tool helping him is Claude Science, a new app recently launched by Anthropic.

On June 30, 2026, Anthropic released Claude Science, positioned as an AI workspace for scientists. (Source: Anthropic official blog)

According to Anthropic, this work would have taken the scientist and his team two years to complete.

Anthropic positions Claude Science not as a smarter scientific model, but as an AI workstation designed for scientists.

Its true breakthrough lies in being the first to break down scientific research into a step-by-step auditable workflow.

Claude Science is now in beta on macOS and Linux, available to Pro, Max, Team, and Enterprise users.

Anyone who has done scientific research understands how tedious it can be:

A project must jump between dozens of databases, each with its own schema and query language;

Files come in a variety of formats, each requiring custom pipelines and viewers to be set up on the spot;

A row of tools sits nearby: PubMed for searching literature, Jupyter for running code, R for statistics, and cluster terminals for submitting tasks…

Constantly switching contexts, the time truly available for thinking about scientific problems is often exhausted by these tasks of moving, stitching, and debugging.

What Claude Science does is bundle these fragmented scenarios into a single execution environment:

Complete all stages—literature analysis, multi-step calculations, chart refinement, and paper finalization—within the same environment, so you never have to interrupt your flow switching tools.

It can run locally on macOS or Linux, connect remotely via SSH, or be mounted on an HPC login node.

Just like you normally use Jupyter, it goes wherever your data is.

Even when it comes to hash rate scheduling, it’s covered.

In the past, folding a protein or running a genomic pipeline on massive datasets required researchers to manually manage the computational tasks, wait in line for cluster resources, monitor whether the job succeeded or failed, and then retrieve the results—eating up half a day in the process.

Claude Science has taken over this process: drafting plans first, consulting you before engaging new resources, allowing you to review or revoke tasks before writing or submitting them, and scaling the analysis from one GPU to hundreds.

Claude Science dispatched an 8-group scVI hyperparameter sweep to the lab's A100 cluster; the right-side notebook shares the same real-time kernel with the agent, enabling live synchronization of variables and state. (Source: Anthropic official blog)

More importantly, sensitive data never leaves the original system; only the context truly needed at each step is sent to Claude.

The field of scientific research is inherently tied to graphics—protein 3D structures, genome browser tracks, and chemical structural formulas are all inherently graphical.

Claude Science, building on this, provides the code used to generate the images and drafts alongside them, and can natively render them as well.

More crucial is reproducibility.

Every time Claude Science generates a chart, it embeds the exact code used to create it, the runtime environment, a plain-language explanation, and the full conversation history directly onto the chart.

On the left, a cell diagram spanning 138 species; on the right, the exact code that generated it is displayed on the same screen—highlighting a single annotation allows the AI agent to modify the image. Every result is reproducible and traceable to its source code. (Source: Anthropic official blog)

A paper often takes more than half a year from submission to publication; months later, when reviewers ask you to rerun a specific figure, you can easily reproduce the entire chain—input, process, and output—on the spot.

Want to change the chart? Just say it—“Remove the gridlines,” “Switch the y-axis to logarithmic”—and the agent will directly modify its own code.

You can fork the conversation at any node to test two ideas simultaneously, without disrupting the original thread.

For the first time, research has been integrated into an auditable workflow, with code, environment, and history enclosed in a closed loop.

Behind Claude Science is not a single agent working alone.

You are facing a coordinating intelligent agent that holds over 60 pre-configured skills and connectors for genomics, single-cell analysis, proteomics, structural biology, and cheminformatics.

With more activity, it can autonomously generate additional agents to divide tasks and instantly invoke expert agents you've created yourself.

The best part is the reviewer agent.

It specifically checks citations and calculations, identifying incorrect citations, numbers without verifiable sources, and figures that don’t match the code—and flags or corrects them automatically.

In the Allen Institute case, the team used an actor-critic pair, with one agent responsible for writing and another专门 evaluating its accuracy and the validity of its citations.

This structure already shows the雏形 of "AI internal peer review."

But one boundary must be clearly stated: the process is always human-in-the-loop.

Before using new resources, it seeks authorization, allowing you to review and revoke every decision. It automates processes—not scientific discovery on your behalf.

It also integrates with NVIDIA’s BioNeMo Agent Toolkit, enabling native connectivity to life science models such as Evo 2, Boltz-2, and OpenFold3.

You can save models, data, and pipelines you trust in your own lab as reusable skills and attach them—future conversations will automatically inherit them.

Claude Science's first focus is on life sciences.

Genomics, single-cell analysis, proteomics, structural biology, and cheminformatics—ready to use out of the box.

It can read scientific literature and query over 60 scientific databases, including UniProt, PDB, Ensembl, ClinVar, ChEMBL, and GEO—no need to learn each one individually.

Claude Science provides pre-configured environments for genomics, single-cell analysis, proteomics, and cheminformatics, backed by 60+ scientific databases. (Source: Anthropic official blog)

Manifold Bio develops tissue-targeted therapies.

They used Claude Science to nominate targets for the latest experiment, evaluating surface expression, trafficking, and safety for each tissue and target, then ranking candidates according to criteria learned from their own data.

Manifold says that ordinary coding assistants can't do this—Claude Science can complete the entire process end-to-end, obtain the right data, make the correct decisions, and incorporate context from past projects.

Here’s an even more advanced example.

An associate professor of epidemiology at the UCSF Brain Tumor Center uses it for molecular epidemiology research on gliomas, analyzing how thousands of small-effect germline variants combine to shape individual susceptibility.

According to Anthropic, Claude Science completed this phylogenetic analysis in approximately one-tenth of the time, and its team independently verified the results, confirming both speed and stability.

However, these 10x speed improvements are currently limited to literature reviews, genomic analysis, and specific pipeline automation—they do not equate to a 10x acceleration of all scientific research.

Meanwhile, the threshold for scientific credibility is being redefined.

In the past, the credibility of a study was measured by peer review and whether it could be replicated by others.

Reproducibility over the long term is one of the biggest pain points in research—once the code is lost or the environment changes, even the authors themselves can no longer reproduce the original results months later.

Each image from Claude Science comes with traceable code, and every result is linked to its environment and history. This barrier of reproducibility may be the first one to be crossed.

In the biotech research sector, the three major players are all competing, but each has a different approach.

Google bets on its proprietary model, OpenAI bets on the scientific intelligence of models, and Anthropic bets on workflows.

Google is directly entering the field with its own proprietary models, such as AlphaFold and AlphaGenome, which others do not have.

OpenAI is following a different path.

In April this year, it launched GPT-Rosalind, a cutting-edge model designed for biological reasoning and drug discovery.

Now, we are taking it a step further by training the model's "scientific judgment."

It has just launched GeneBench-Pro, a test designed to evaluate whether models can make judgments like computational biologists: 129 questions spanning genomics and population genetics to clinical diagnostics, specifically assessing the ability to determine “whether the data supports the question” and “when to backtrack and start over.”

The strongest GPT-5.6 Sol achieved 28.7%; with Pro mode enabled, it reached 31.5%. Earlier generations of GPT-5 scored less than 5%.

OpenAI itself says that, at this rate, it could be overwhelmed by the end of the year.

Even the strongest models can solve less than one-third. The part they cannot solve is precisely where human scientists stand.

The AI shortcomings exposed by GeneBench-Pro are also clear:

The model can get things started, but it can’t close the final loop—decisions like whether to remove a set of anomalous data or how to adjust the approach after a hypothesis is overturned still require scientists to make the final call.

Claude Science also does not bypass this point: proposals are reviewed by humans, and every decision remains subject to human reversal. It automates the process, but decision-making authority is not delegated to the model—humans remain in the loop.

For scientists like Lecoq, whether a review can be replicated and still hold up months later is far more important than a fraction of a percentage point on a ranking list.

Claude Science is betting on making AI research a reality in everyday laboratory settings.

Reference materials:

https://www.anthropic.com/news/claude-science-ai-workbench

https://openai.com/index/introducing-genebench-pro/

This article is from the WeChat public account "New Intelligence Yuan," authored by ASI Revelation.