Sam Altman lauds Kosmos, the AI scientist who finds new disease clues and does months of work in a day

OpenAI chief Sam Altman has praised a major leap in artificial intelligence research after Future House unveiled Kosmos, a next-generation AI Scientist designed to speed up scientific discovery. In a post on Sunday, Altman called the development “exciting” and said systems like this would become some of the most important impacts of AI in the years ahead.

According to a detailed blog post published last week, Kosmos represents a significant upgrade over Future House’s previous AI Scientist, Robin, and is now being managed by Edison Scientific, a new commercial spin-off that will operate the platform while maintaining a generous free tier for academics.

What Kosmos aims to solve

The blog post notes that traditional AI models struggle to handle huge amounts of scientific information because they can only keep a limited amount of context in memory at a time. This makes it difficult for them to follow long, complex chains of reasoning.

Kosmos tries to overcome this by using structured world models. This approach lets it draw on information collected across hundreds of agent runs while staying focused on a single research goal, even when working through tens of millions of tokens.

A typical Kosmos run involves reading around 1,500 scientific papers and executing more than 42,000 lines of analysis code, far more than previous systems used in research, adds the report.

Early users estimate that one Kosmos run can complete work that would normally take a human scientist around six months. Internal testing found that almost 80% of Kosmos’ conclusions were accurate.

Seven discoveries across multiple fields

Future House published seven discoveries made by Kosmos during testing, covering neuroscience, materials science, genetics and ageing research.

Reproducing earlier human findings

Kosmos successfully matched three scientific findings previously made by researchers:

• It confirmed unpublished work showing that nucleotide metabolism changes sharply in the brains of hypothermic mice.
• It independently reproduced a materials science finding that absolute humidity during thermal annealing is the key factor affecting perovskite solar cell performance, including the failure threshold around 60 g per cubic metre.
• It identified the same mathematical rules for neuronal connectivity across species as reported in earlier academic work.

New contributions to science

Kosmos also generated four novel insights:

• It suggested that higher levels of the antioxidant enzyme SOD2 may reduce heart fibrosis in humans, supporting findings previously seen only in mice.
• It proposed a new molecular explanation for how a genetic variant may lower the risk of Type 2 diabetes.
• It developed a new method for analysing how tau proteins accumulate in the brain in Alzheimer’s disease.
• In a large scale analysis of ageing in neurons, it found that certain neurons in the entorhinal cortex lose expression of flippase genes with age. This could expose “eat me” signals on cells and make them more vulnerable to attack by microglia. The pattern was later observed in human Alzheimer’s datasets, adding weight to the claim.

Edison Scientific said several of these findings are now being validated in wet lab experiments.

A key feature of Kosmos is traceability. Each conclusion can be linked back to specific lines of code or specific scientific passages that informed it. The company says this ensures all reports remain auditable and avoids the “black box” problem common in many AI tools.

How much does it cost?

Kosmos is available now, priced at 200 credits per run, equivalent to $200. Academics will receive some free usage, and early subscribers can lock in the current price before it rises.

The team stresses that Kosmos is not a chatbot. Instead, it behaves more like a scientific reagent kit that researchers deploy for high-value questions. It can also produce dead ends or follow statistically interesting but scientifically irrelevant patterns, meaning multiple runs may be needed for certain projects.

Why a single run may equal six months of work

The claim that Kosmos can deliver the equivalent of six months of scientific labour surprised even its creators. The estimate came from beta users, who reviewed Kosmos’ outputs and compared them against how long it would take them to reach the same conclusions manually.

The team also notes that three of Kosmos’ reproduced discoveries originally took human researchers roughly four months to complete. When considering the number of papers read and analysis steps performed per run, Edison Scientific argues that the time saved aligns with several months of full-time research.