Google DeepMind's Science Skills Pushes Gemini 3 Flash From 49% to 93% Accuracy
Google DeepMind open-sources Science Skills, a 34-tool bundle that cuts token usage in half and boosts agent reliability for biology and life science workflows
Takeaways
- Open-sourced: Google DeepMind released Science Skills on GitHub (Apache 2.0), a 34-tool bundle for agentic scientific workflows.
- Massive reliability boost: Gemini 3 Flash task success jumped from 49% to 93%; Pro model from 67% to 91% on internal benchmarks.
- 2x token efficiency: Skills cut average token usage by up to 2.04x by replacing generic web search with targeted API calls.
- Flash matches Pro: With Science Skills, the cheaper Flash model matches the reliability of Pro without skills, a significant cost implication.
- 34 databases covered: Integrates AlphaFold, UniProt, gnomAD, ClinVar, PubMed, ChEMBL, PDB, GTEx, and 26+ more life science sources.
- Part of Gemini for Science: Announced at Google I/O alongside Literature Insights, Hypothesis Generation, and Computational Discovery tools.
Agentic AI has been making inroads into software engineering for a while, but science has proven harder to crack. The problem is not that language models lack general knowledge about biology or chemistry. It's that when an agent needs to look up a specific protein structure, query a clinical variant database, or ground a claim in a published paper, it has no reliable way to do so. It falls back on web search, hallucinates citations, and burns through tokens trying to recover. Google DeepMind's Science Skills is a direct attack on that problem.
What just shipped
Google Antigravity is Google's agentic development platform, built on the Gemini model. The Science Skills bundle augments it with science-specific instructions in the form of agent skills, strengthening Antigravity for common scientific workflows. The toolkit is now publicly available on GitHub under an Apache 2.0 license and free to use.
As part of the broader Gemini for Science initiative, Science Skills integrates insights from over 30 major life science databases and tools including UniProt, AlphaFold Database, AlphaGenome API, and InterPro. The initial focus is biology and life sciences, covering four main domains:
- Genomics, Transcriptomics, and Regulatory Biology (11 skills): variant analysis via ClinVar, dbSNP, gnomAD; gene expression via GTEx and Ensembl; regulatory elements via ENCODE and JASPAR
- Proteomics and Structural Biology (10 skills): AlphaFold Database, PDB, UniProt, Foldseek structural search, PyMOL visualization, protein sequence alignment
- Cheminformatics and Clinical Translation (5 skills): ChEMBL, ClinicalTrials.gov, OpenFDA, OpenTargets, PubChem
- Scientific Literature, Pathways, and Ontologies (8 skills): PubMed, arXiv, bioRxiv, Europe PMC, OpenAlex, Reactome, QuickGO
There is also a Workflow Automation skill that lets you teach the agent your own custom workflows. A user can "handhold" an agent through their intended workflow and then use that interaction as a template to create their own reusable skill tailored to their needs.
The architecture: skills as structured instructions
The term "skill" here has a specific technical meaning. Skills point to key resources by simple text files and scripts which enable the agent to efficiently deliver the right information for models to utilize in grounding their statements and reasoning. They follow the Agent Skills standard, a cross-vendor specification that lets you customize agents through a layered set of instructions, code, and data.
Each skill directory contains three things:
- A
SKILL.mdfile with YAML frontmatter and step-by-step instructions (think of it as a system prompt for a specific task) - A
scripts/folder with helper Python scripts that wrap the relevant APIs - Optional
references/documentation
You install the whole bundle with a single command:
npx skill add google-deepmind/science-skills/Existing Antigravity users can update to the latest version, then navigate to Settings → Customizations → Build with Google Plugins to download the Science plugin. The first time you trigger a Science Skill, the agent will ask for approval and install uv, then proceed to respond to your query. Some skills like AlphaGenome require an API key; others like ClinVar work without one but unlock higher rate limits with a key.
The numbers that matter
The team published a technical report with rigorous benchmarks across two model sizes. The results are striking.
On an internal benchmark of 67 expert-validated tasks, Gemini 3 Flash went from 49% task success to 93% with Science Skills. Gemini 3.1 Pro went from 67% to 91%. The gap between the cheaper Flash model and the more expensive Pro model essentially disappeared once skills were added.
| Model | Without Skills | With Skills | Token Efficiency Gain |
|---|---|---|---|
| Gemini 3 Flash | 49% reliability | 93% reliability | 2.04x fewer tokens |
| Gemini 3.1 Pro | 67% reliability | 91% reliability | 1.62x fewer tokens |
Token consumption dropped by between 1.5x and 2x across both models, in addition to the performance gains. On the external BioReason benchmark, which tests multi-step variant effect reasoning across 2,106 tasks, Science Skills improved accuracy from 41.4% to 60.9% on coding variant prediction, and from 46.6% to 81.6% on non-SNV variant prediction.
Why the efficiency gains are real
The token savings are not just a side effect of the agent succeeding faster. They reflect a qualitative change in how the agent reasons. Without Science Skills, Antigravity uses more steps, makes errors along the way, and often reverts to generic web search to find specific information. With Science Skills, the overall number of steps is reduced and reasoning is grounded in reference data.
The agent traces in the paper make this concrete. For a simple query about the protein components of EGFR target CHEMBL203, the agent without skills fired off 11 separate search and grep operations before arriving at the answer. With the chembl-database skill, it ran a single targeted API call and got the answer in 6 steps. For a harder task about ACE2 protein expression in specific tissues, the agent without skills scraped Wikipedia, hit HTTP 400 errors on the UniProt API, and still returned a wrong conclusion. With the human-protein-atlas-database skill, it resolved the Ensembl ID, queried the right endpoint, and returned the correct answer in 4 steps.
What it's good at, and where it falls short
The bundle excels at exactly the kind of multi-step lookup chains that are common in life science research: starting from a gene name, pulling variant frequencies from gnomAD, checking clinical significance in ClinVar, fetching the protein structure from AlphaFold, and grounding everything in a PubMed citation. Science Skills enables researchers to complete workflows such as structural bioinformatics and genomic analysis in minutes rather than hours.
The limitations are worth knowing upfront:
- Coverage is narrow: The initial bundle focuses on biology and life sciences. Chemistry, physics, materials science, and climate science are not yet covered.
- Long workflows are not tested: The current skills are intended to help with workflows that consist of dozens of steps, last a few hours, and operate perhaps half a dozen different skills within a single task. Evaluation flows do not include particularly long-running or complex multi-step tasks.
- Environment variability: The team adopted the
uvpackage manager to enforce reproducibility, which adds constraints and can impact the user experience on some desktop setups. - Only works with Antigravity: The skills are designed for Google's own agentic platform. Porting them to other agent frameworks (LangChain, AutoGen, etc.) would require adaptation.
The bigger picture
Google launched Gemini for Science at Google I/O, announced by James Manyika, Senior Vice President at Google, and Pushmeet Kohli, VP at Google DeepMind and Chief Scientist at Google Cloud. Science Skills is one component of a larger push that also includes Literature Insights (powered by NotebookLM), Hypothesis Generation (Co-Scientist), and Computational Discovery (AlphaEvolve).
DeepMind teams tested Science Skills on a rare genetic disease caused by AK2 mutations, which underlie reticular dysgenesis, a severe combined immunodeficiency. Using Science Skills, researchers obtained a complex structural analysis of AK2 and disease-linked variants significantly faster than with conventional methods, supporting new mechanistic insights into how AK2 loss-of-function disrupts energy metabolism, hematopoiesis, and sensory hair cell development.
Google is also collaborating with over 100 institutions, including Stanford University on liver fibrosis, Imperial College London on antimicrobial resistance, and a multi-year effort with The Crick Institute, to validate these new systems and tools.
The deeper implication for the field is about model sizing. The conventional assumption is that harder tasks require more capable (and more expensive) models. Science Skills challenges that directly: a Flash-class model with the right domain scaffolding can match a Pro-class model operating blind. That is a meaningful result for anyone building research pipelines where cost and latency matter.