Claude Science Ships as a Workbench With User-Billed Modal Compute

Anthropic launched Claude Science on June 30, 2026: a standalone beta workbench, not a new model, for Pro, Max, Team, and Enterprise on macOS and Linux. The app ships with 60+ curated scientific skills and connectors. GPU jobs that outgrow a laptop go to the user's own Modal workspace. Modal bills those jobs. Anthropic does not.

MIT Technology Review framed the launch as Anthropic's newest flagship product, ranked with Claude Code and Claude Cowork. The Anthropic post is more specific: an app that runs analyses, writes auditable artifacts, and submits compute to a laptop, an HPC cluster over SSH, or a Modal account, scaling a job from a single GPU to hundreds.

Workbench, not a biology model

Claude Science is a research environment. It uses the Claude models already on the subscriber's plan. Anthropic's product page states that directly: "Claude Science is a public beta app, not a model." What changed is the harness around those models: scientific databases, specialist agents, a reviewer that flags bad citations and untraceable numbers, and compute that stays on the lab's infrastructure.

The June 30 post lists the day-one domains: genomics, single-cell, proteomics, structural biology, and cheminformatics. Specialist agents query sources such as UniProt, PDB, Ensembl, Reactome, ClinVar, ChEMBL, and GEO. Native renderers cover 3D proteins, genome tracks, and chemical structures. Every figure ships with the code, environment, and message history that produced it.

NVIDIA's BioNeMo Agent Toolkit is wired in as skills. The named models are Evo 2, Boltz-2, and OpenFold3.

Modal is the user's bill, not Anthropic's

Default execution is local: a sandboxed environment on the researcher's machine. When a job needs a GPU or more CPUs than that machine has, Claude Science routes it to a connected Modal workspace. Modal's June 30 post is explicit: the researcher connects their own workspace; GPU and concurrent-CPU workloads land in Modal sandboxes; the researcher stays in Claude.

Claude Science compute docs match that billing split in one paragraph: "You connect your account, jobs run on it, and Modal bills you directly. Anthropic doesn't provide or bill compute and never sees a payment method." Each job shows a confirmation card with the machine spec, per-second billing, and a maximum billable time. There is no spend ceiling inside Claude Science. Closing the app does not cancel a running Modal job.

Modal's post says a new account includes $30 of free compute at signup. Anthropic's launch post is a different pot: up to 50 AI for Science projects, up to $30,000 in Claude credits, plus up to $2,000 of Modal compute for select projects. Applications closed July 15, 2026. Awards were due by July 31. Project window: September 1 to December 1, 2026.

"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, Chai Discovery — Modal blog, June 30, 2026

Named labs, then the protein-binder dataset

Anthropic named three early users. Manifold Bio used the workbench to nominate tissue-targeting binder targets, scoring surface expression, trafficking, and safety against its own program data. Jérôme Lecoq at the Allen Institute built a multi-agent review pipeline of about 20 custom skills; the team reports about 10 long-form reviews, many over 100 pages, with citations checked by reviewer agents. Stephen Francis at the UCSF Brain Tumor Center used it on molecular epidemiology of glioma and said germline workups that used to take the full cycle now run in roughly one-tenth the time, with the lab independently validating the outputs.

The prompts that produced Anthropic's protein-binder campaign assume this harness. The Hugging Face dataset Anthropic/claude-protein-binder-design ships 1,440 de novo miniprotein binders (50 to 120 residues) against 16 targets: 900 from Mythos Preview, 540 from Opus 4.8. Two CROs measured them: Adaptyv Bio and Twist Bioscience. Mature GDF-8 (120 designs) was inconclusive. Across the other 15 targets, 354 of 1,320 designs are binders by the release's two-vendor assessment.

The prompt README is the product link. The campaigns assume the Claude Science agent harness: host.compute, submit_gate, and Modal sandboxes tagged claude-science-project, plus an operator-funded Modal account. That is prior coverage, not a second wet-lab scoop: see Ultrathink's protein-binder article.

What the launch actually ships

Claude Science is the app. Modal is the burst GPU bill the user already has, or can open for $30. The Hugging Face release is the campaign that was written for that stack. The model in the loop is the same Claude the lab already pays for.