Anthropic integration with Modal brings scalable compute to Claude Science

All posts

Back

News

June 30, 2026•6 minute read

Anthropic integration with Modal brings scalable compute to Claude Science

Today Anthropic launches Claude Science, Anthropic’s AI workbench for scientists, where life sciences researchers can run computational workloads directly from a conversation with Claude - from data processing pipelines to structure prediction models to molecule design campaigns.

The most demanding problems in the life sciences need more than the most capable AI models. They need world-class infrastructure to run on. Running inference against protein language models and screening libraries of millions of compounds: this is compute-heavy work that has to scale on demand and vanish when it's done. We're proud to integrate with Claude Science, Anthropic’s AI workbench for scientists, for that compute.

Computational biology has a toolchain problem

Most life sciences pipelines span local machines, HPC clusters, cloud VMs, and a rotating cast of bioinformatics tools.

A typical workflow: provision the right kind of resources or submit to a job a queue, ssh into the servers, set up environments, run jobs, explore and analyze results, and repeat for every workload.

Claude Science pulls that work into one place. Researchers run code directly inside the conversation. By default that execution happens in a sandboxed environment on the researcher's machine, which is plenty for early-stage analysis. But computational biology workflows have a habit of starting small and turning demanding. That's where Modal comes in, giving you access to more capacity automatically when you need it, without having to change anything else.

When that happens in Claude Science, researchers can connect their own Modal workspace. Workloads that require GPUs or many concurrent CPUs route to Modal sandboxes automatically. The researcher stays in Claude. Modal handles the compute: bursty, heterogeneous workloads that outgrow any fixed cluster or single GPU type.

What Modal makes possible

Fan out. Variant analysis, virtual screening, and sequence annotation are all highly parallel workloads. A virtual screen run against a large compound library that would take hours sequentially fans out across hundreds of Modal containers and comes back in minutes. Modal handles the scheduling; Claude writes Python.

GPU access per step. Biology pipelines are rarely all-GPU or all-CPU. A typical workflow might do alignment and preprocessing on CPU, then hand off to ESMFold2 or a similar model for structure prediction on GPU. With Modal, you request a GPU for that one function, not for the whole pipeline.

Shared storage across jobs. Moving around things like sequencing datasets, protein databases, and model checkpoints between pipeline steps is expensive. With Modal Volumes, it's done once and available across every job, whether you're running one query or thousands.

Reproducible environments. Modal Images make it easy to define and manage dependencies for each type of compute job, so environments are concise and consistent across runs. No surprises when someone else on the team tries to reproduce a result.

Biology teams are already running this way on Modal

Chai Discovery builds ML-driven drug discovery pipelines that chain together protein embeddings, multiple sequence alignments, and antibody design models, each step running on whatever hardware matches its compute shape.

“We have a lot of different models that give us different layers of insights on these proteins, and being able to run them all on the hardware that makes sense is what makes the product possible.”

— Kevin Wu, Machine Learning Researcher

The workloads are bursty by nature: quiet one day, hundreds of GPUs the next.

“Sometimes we spin up hundreds of GPUs at a time, and the fact it's up in a few minutes without onerous configurations or dashboards whenever we need to is kind of a miracle.”

— Kevin Wu, Machine Learning Researcher

Claude Science puts that same elastic, reproducible infrastructure within reach for researchers who've never touched MLops.

What it looks like in Claude Science

The smallest version of the pattern is a single prompt. A researcher types "Use a Modal GPU to predict the 3D structure of GFP, then show me the predicted structure and its confidence scores." Claude writes the folding code, runs it on the researcher's connected Modal workspace, and returns the structure with per-residue confidence, all inside the conversation.

Engineering an enzyme across multiple models. A researcher uploads an enzyme sequence and asks Claude to prioritize candidate activity-enhancing mutations, scored across several state-of-the-art methods. Claude characterizes the protein, folds it with ESMFold, then scores every possible single mutation with ESM-2, ProteinMPNN, and ESM-IF: three different model families. Modal gives each step the hardware it needs; Claude ranks the consensus candidates and renders the top mutations on the structure.

Designing a genome-wide CRISPR knockout screen. "Design a knockout screen targeting every human kinase: select guides, predict off-targets, plan sequencing depth." The off-target search is the expensive step: every candidate guide has to be checked against the whole human genome for near-matches. Sequentially, that crawls. Fanned out across hundreds of Modal containers, thousands of guides are screened in parallel and come back in minutes. Claude aggregates the hits, flags the risky guides, and assembles a construction-ready library.

Single-cell analysis at dataset scale. "Find a study comparing tumor biopsies before and after immunotherapy, then identify which immune populations expand or contract with treatment." Claude locates the dataset, builds the single-cell object, clusters and annotates the cell types, and identifies which immune populations expand or contract with treatment.

Getting started

Claude Science launches June 30th. To use Modal as your compute backend:

Download the Claude Science app at claude.com/science.

In Settings, navigate to the Compute configuration section.

Connect to your Modal workspace. If you don't have a Modal account yet, create one at modal.com/signup — it takes a few minutes and includes $30 in free compute to start.

Once connected, compute workloads will run on Modal automatically.

Read Anthropic's full launch announcement.

Compute support for life sciences departments

Claude Science is built for researchers at universities and institutions, not just commercial biotech. We're making Modal compute available to life sciences departments through this launch.

We're committing up to $100,000 in compute to support Anthropic’s AI for Science Claude Science Cohort, with allocations of $500–$2,000 per project. Applications are open through July 15, 2026, with award notifications sent out by July 31. Projects will run from September 1 to December 1, 2026. Apply here.