$ cat wiki/models/co-scientist.md

Co-Scientist (Google DeepMind)

modelupdated 2026-07-21created 2026-05-22

Spec

Attribute	Value
Developer	Google DeepMind
Released	2026-05-21 (Nature paper + experimental researcher rollout)
Announced	2026-05-19 (Google I/O 2026, as part of Gemini for Science)
Context window	unknown
Pricing	unknown
License	unknown
Availability	Gemini for Science program — individual researcher request (experimental rollout as of 2026-05-21)
Type	Multi-agent AI system for scientific hypothesis generation and experiment design
Underlying model	Gemini (exact version unspecified; Gemini for Science suite)
Paper	Published in Nature (2026-05-21)

Scientific hypothesis generation: Given a research question or grant proposal summary, generates and ranks novel hypotheses
Infectious disease research: Identifying molecular switches in zoonotic disease (Ebola, HIV, flu, Covid-19)
Molecular biology: Amino acid-level discovery in protein function
Other Co-Scientist applications: aging research (Cambridge), ALS (University of Edinburgh), liver disease

Cambridge infectious disease case: Identified a protein not on the researcher's radar → path to amino acid discovery in 6 months vs. 2-3 years normally (~4× speedup)
Hypothesis generation goes "beyond human prior knowledge" — surfacing candidates the human expert hadn't considered
Published in Nature: validates scientific rigor (peer-reviewed)

First major AI system for hypothesis generation to receive a Nature publication
Moves from research demo → production tool for working scientists
Key data point for the "AI-accelerated R&D" thesis (one of Anthropic Institute's 4 pillars; OpenAI's "Year of Science" framing)
Complements AlphaFold (structure prediction) and AlphaEvolve (algorithm design) in DeepMind's AI-for-science stack
If the 4× speedup generalizes across biology, it changes the economics of wet lab research

System	Focus	Access	Status
Co-Scientist	Hypothesis generation, biology/science	Request-based (Gemini for Science)	Nature paper, experimental rollout
AlphaFold 3	Protein structure prediction	Open access (academic)	Production
AlphaEvolve	Algorithm design (coding)	Internal + commercial partners	Production
OpenAI GPT-Rosalind	Life sciences reasoning	API	Production

Does the 4× speedup generalize beyond molecular biology to physics, chemistry, social science?
What's the false positive rate on Co-Scientist hypotheses — how often does the ranked hypothesis pan out?
Will Co-Scientist move beyond request-based access to self-service?