Anthropic Could Have Shipped a Bigger Model. It Shipped 60 Skills Instead.

On June 30, Anthropic launched Claude Science, a workbench built on the same local-kernel bet MegaBrain Science makes. It didn't ship a smarter model. It shipped 60-plus curated skills and connectors for genomics, proteomics, and structural biology. Three papers from the same ten days explain why that's the smarter bet: a frontier agent given free rein on a real research problem beats the published state of the art on 17.8% of tasks. Give the identical model pre-built domain skills, and completion on a biomedical benchmark rises from 57.1% to 100%. Same model. Different toolkit. Different outcome.

The launch: Claude Science

Anthropic's Claude Science went into beta on June 30 for Pro, Max, Team, and Enterprise users on macOS and Linux. It pairs a local kernel with "over 60 curated skills and connectors" pre-configured for genomics, single-cell, proteomics, structural biology, and cheminformatics, and produces reproducible artifacts with auditable histories. Anthropic is also funding up to 50 "AI for Science" projects with up to $30,000 in Claude credits (applications close July 15).

It's the clearest validation yet, from the field's largest lab, that a local kernel plus deep domain integration — not a bigger model — is the right shape for a research workbench. See our full, source-by-source comparison at /product/megabrain-science/compare.

The number behind the launch: 17.8%

NatureBench(arXiv 2606.24530, June 23) built 90 tasks directly from peer-reviewed Nature-family papers and ran ten agent configurations against them. The best configuration beat the paper's own published result on just 17.8% of tasks — and the paper's failure analysis pins the gap on "wrong method choice and insufficient compute budget," not broken reasoning.

The number that explains the fix: 57.1% → 100%

NVIDIA's BioNeMo Agent Toolkit (June 23) packages protein-structure prediction, molecular generation, docking, and sequence analysis as MCP-documented "skills." The same reference agent — GPT-5.5 fast via Codex CLI — went from 57.1% to 100% task completion once it had those skills to call, with 2x more passing assertions per token consumed. Nothing about the underlying model changed.

Already in production: Talos

Microsoft Research's Talos (June 24) is an open-source agent that continuously reanalyzes genomic data for undiagnosed rare-disease patients as new evidence is published. Deployed across 4,735 individuals over 29 monthly cycles, it surfaced 241 new diagnoses (a 5.1% additional yield), with a median 32 days from a new piece of evidence to a diagnosis — and 59% of those new diagnoses weren't yet curated in OMIM.

Also this week

Elicit shipped a more powerful Research Agent (June 30) that now ingests gene sequences, omics datasets, microscopy images, and lab file formats like PDB and FCS directly. Sakana AI published three research posts in ten days, including CoffeeBench, a 90-day multi-agent supply-chain simulation where Claude Haiku 4.5 was the one model to post a loss, and Sheaf-ADMM, a coordination method that hit a 93% solve rate on multi-agent Sudoku versus an 11% baseline. And Edison Scientific, maker of the Kosmos platform, partnered with Population Health Partners (June 29) to apply it to population-scale disease biotech.

What this means for local-first, reproducible science

The frontier labs are converging on the same answer to the model question: the gap between a chatbot and a scientist isn't parameters, it's tooling and integration. That's exactly the gap a locally-run, deeply-integrated kernel is built to close — we dig into the evidence in today's deep dive.