Up to 6 months of research per run (edisonscientific.com)

FutureHouse today launched Kosmos, a next‑generation “AI Scientist” (now operated by spinout Edison Scientific) that uses structured world models to sustain long-horizon, research‑grade reasoning. Unlike prior agents limited by LLM context windows, Kosmos aggregates information from hundreds of agent trajectories and maintains coherence over tens of millions of tokens; a single run ingests ~1,500 papers and executes ~42,000 lines of analysis code. Beta metrics report 79.4% of conclusions as accurate and users estimated a typical deep run delivers the equivalent of ~6 months of PhD/postdoc effort. Kosmos is live on the platform ($200/run with academic free tier), designed for traceability (every conclusion links to code or literature), and is intended as a research reagent rather than a chatbot. The team reports seven validated discoveries across neuroscience, materials and statistical genetics: three independently recapitulated human findings (metabolomics in hypothermic mouse brain, a perovskite humidity “fatal filter” at ~60 g/m³, and cross‑species neuronal connectivity rules) and four novel contributions (MR evidence implicating SOD2 in reduced myocardial fibrosis, a SNP mechanism lowering Type‑2 diabetes risk, a proteomics method for ordering tau‑accumulation events, and a clinically supported finding that age‑related flippase loss in entorhinal neurons may expose “eat‑me” signals linked to early Alzheimer’s pathology). Important caveats: the 6‑month estimate is beta‑user based and Kosmos can chase spurious leads, so outputs require human review and wet‑lab validation; nonetheless, the system points to practical inference‑time scaling laws for AI‑accelerated discovery.