KKosmos by Edison Scientific

A single Kosmos run reads about 1,500 scientific papers and writes an average of 42,000 lines of analysis code, then hands back a research report with citations and the code behind every claim. Independent scientists who checked those reports found 79.4% of statements accurate. That combination of scale and a published error rate is what sets Kosmos apart: it does more autonomous work than most research agents, and Edison Scientific is open about how often it gets things wrong.

Kosmos suits academic labs and biotech or pharma research teams with real datasets and the expertise to check results. Cost per run is the main point to plan for: a Kosmos run costs 200 credits, so a 650-credit month buys three, and every report is designed for expert checking. Academics get 650 credits a month free, and the paid plan was $200 a month as of December 2025. We rate Kosmos 8.3/10 on our scale. We have not used the product; this review is based on Edison’s documentation and announcements and on news coverage.

Run Statistics, Funding and Pharma Deals Behind Kosmos

Edison has published more hard numbers about what its agent does in a run than most research tools, along with funding and pharma deals that show who is betting on it.

The accuracy number deserves the most attention. Roughly one statement in five did not hold up under independent review, which is a strong result for an autonomous system and a clear reason to treat every report as a draft. Plan for a scientist to check the claims that matter before anyone acts on them.

The Incyte deal points to where the product is heading. Incyte expects its own data to help train Kosmos, so the system learns from a sponsor’s data rather than just analyzing it. That is attractive for large companies with deep archives, and it raises questions about data rights that belong in the contract.

The Population Health Partners deal goes further still. Kosmos is embedded across ideation, company creation and drug development, starting with a clinical-stage asset PHP says could affect 10% of the world’s population, with cardiovascular, inflammatory lung and gastrointestinal disease named as later focus areas. Incyte’s global head of R&D, Pablo Cagnoni, framed his company’s deal as a way to get more value from its data by having AI guide experimental design, while Edison CEO Sam Rodriques described the goal as a system that keeps learning from data rather than only analyzing it.

An Objective and a Dataset In, a Cited Report Out

Kosmos runs on Edison’s web platform. According to Edison’s documentation, you give it a research objective and one or more datasets, and it reads literature, writes and runs analysis code, generates hypotheses and produces a cited report.

Inside a Run

• Read: Kosmos searches and reads the literature relevant to the objective, about 1,500 papers in a typical run.
• Analyze: it writes and executes analysis code on the uploaded data, averaging 42,000 lines per run.
• Hypothesize: it forms hypotheses and tests them against data and literature, repeating over many cycles.
• Report: it assembles a report with citations and code, and Edison says every conclusion is fully auditable.

Collaborators estimated that a 20-cycle run equaled about six months of their own research time, and Kosmos reproduced findings that had taken human researchers about four months. Edison calls these estimates rough.

From FutureHouse to Edison

Kosmos grew out of FutureHouse, a nonprofit research lab whose platform launched in May 2025 with earlier agents named Crow, Falcon, Phoenix and Owl, offered through a developer platform and API. Demand from pharmaceutical and biotech companies outgrew what a nonprofit could support, so Edison Scientific was spun out to handle payments, customer support, enterprise deployment and private investment, while FutureHouse kept its focus on basic research. Kosmos was Edison’s first product.

The Other Agents on the Platform

• Literature: answers cited questions or runs deep reviews across more than 175 million papers, trials and patents.
• Analysis: processes experimental data such as flow cytometry and RNA-seq and produces statistics and publication-ready figures.
• Precedent: checks whether a research idea has been tried before and where the gaps are.
• Molecules: a chemistry agent for molecular design and analysis.

Edison also releases PaperQA, an open-source literature agent, which researchers can run on their own.

Free Academic Credits and Preparing Clean Datasets

• Sign up: create an account on the Edison platform; academics and students qualify for 650 free credits a month.
• Prepare data: structured datasets with clear labels give Kosmos the best chance, since Edison says it works best with structured data and established methods.
• Write the objective: state the research question precisely, because the run follows it for many cycles.
• Review the report: check citations, rerun the main analyses and validate the findings that matter.

Scientists Still Pick the Question and Judge the Hypotheses

Edison says Kosmos is not a replacement for scientists and that its findings should guide investigation rather than settle it. People choose the question, supply good data, judge which hypotheses deserve lab time and catch the roughly 20% of statements that independent reviewers found inaccurate.

Credit Prices Edison Set in December 2025

Edison prices its platform in credits. The figures below were published by Edison in December 2025, when it said the $200 subscription might be phased out at its next major product update.

Three Free Runs a Month for a Graduate Student

These scenarios use the published December 2025 credit prices. Usage levels are our own illustrative assumptions.

• A graduate student: 650 free academic credits cover three Kosmos runs (600 credits), leaving 50 credits for about 25 to 50 runs of the 1 to 2 credit agents.
• A biotech scientist on the $200 plan: three Kosmos runs a month use 600 of 650 credits, about $66.67 per run, while Literature and Analysis runs in the web app cost nothing extra.
• A five-person discovery team: one subscription each comes to $1,000 a month, or $12,000 a year, for up to 15 Kosmos runs a month across the team.

Other costs are mostly human: scientist time to review long reports, rerun code and validate claims. API use is billed separately, and pricing may change if Edison retires the subscription.

Kosmos Compared With Elicit, Causaly Agentic Research and K Pro

Research agents for life sciences range from literature tools to full AI scientists. Kosmos sits at the autonomous end because it analyzes your data as well as the papers.

Elicit is the safer pick for evidence reviews where every claim must trace to a paper. Kosmos goes further by analyzing your own data, which is also where more can go wrong. For more choices, see our life sciences agents page and research agents page.

Academic Labs and Biotech Teams With New Datasets

• Academic labs: free monthly credits make three full runs a month possible at no cost.
• Biotech teams exploring a new dataset: Kosmos can work through literature and analysis in one run and suggest hypotheses worth testing.
• Computational biologists: reports include the code, so results can be rerun and checked.
• Company builders and venture studios: the Population Health Partners deal uses Kosmos across ideation, company creation and drug development, not only for single analyses.
• Pharma R&D groups with large archives: the Incyte deal shows Edison will embed Kosmos with a sponsor’s translational and clinical data.

Groups With No Reviewers or Analyses Bound for Regulators

• Teams without scientific reviewers: with roughly one statement in five judged inaccurate, reports are best checked by an expert before use.
• Regulated analyses: outputs need validation before they go anywhere near a submission.
• Teams that need lab system connectors: Kosmos works from uploaded datasets, so groups that want agents acting inside an ELN or LIMS will prefer Benchling AI.
• Literature-only users: a tool such as Elicit covers evidence review without the cost of full runs.

Planning Around a 79.4% Accuracy Rate

• Accuracy: the 79.4% figure is good for an autonomous system and still leaves a meaningful error rate.
• Input quality: Edison says Kosmos depends on good data and works best with structured datasets and established methods.
• Credit use: at 200 credits per run, Kosmos uses credits about 100 to 200 times faster than the other agents.
• Changing plans: Edison said the $200 subscription may be phased out at its next major update.
• Data rights: deals like Incyte’s involve training Kosmos on sponsor data, so clarify ownership and reuse terms.
• Rough effort estimates: the six-months-per-run comparison comes from collaborators and Edison calls it rough, so do not plan headcount around it.

The FutureHouse Spinout and a $70 Million Seed Round

• On November 7, 2025, HPCwire reported the spinout of Edison Scientific from FutureHouse and the launch of Kosmos.
• On December 18, 2025, Edison announced a $70 million seed round led by Triatomic Capital and Spark Capital, along with free credits for academics.
• In May 2026, Fierce Biotech reported that Incyte will embed Kosmos across its discovery and development data.
• On June 29, 2026, Edison and Population Health Partners announced a partnership to build new biotech companies with Kosmos, starting with a clinical-stage asset.

Worthwhile When Good Data Meets People Who Can Check It

For research teams with good data and people who can check the output, yes. Kosmos scores 8.3/10 on our scale, with a high autonomy score, and more integrations and enterprise controls would lift it further. Reading about 1,500 papers and writing tens of thousands of lines of code in one run is work no literature assistant attempts, and the free academic tier makes it easy to try on a real problem.

Treat each report as a well-researched draft. With about one statement in five judged inaccurate, the value comes from faster hypotheses and analysis, checked by a scientist. Companies should also settle data and training terms before uploading proprietary datasets. A sensible trial is to give Kosmos a question your team has already answered, compare its report with your own findings and use that result to decide how much review each future run needs.

Run Costs, Academic Access and Accuracy of Kosmos Reports

How Much Does a Kosmos Run Cost?

A Kosmos run costs 200 credits. On the $200 monthly plan with 650 credits, three runs work out to about $66.67 each, based on Edison’s December 2025 pricing.

Is Kosmos Free for Academics?

Yes. Edison gives academics and students 650 credits a month, enough for three Kosmos runs, and said this offer has no end date.

How Accurate Are Kosmos Reports?

Independent scientists found 79.4% of statements in Kosmos reports accurate. Edison says the reports are fully auditable, with citations and code, so experts can check each claim.

Who Makes Kosmos and How Is It Funded?

Edison Scientific, a San Francisco company spun out of the nonprofit FutureHouse in 2025, builds Kosmos. It raised a $70 million seed round in December 2025 led by Triatomic Capital and Spark Capital.

How Is Kosmos Different From Elicit?

Elicit focuses on finding and summarizing evidence from papers and trials. Kosmos also writes and runs analysis code on your own datasets and generates hypotheses over many cycles, which costs more per run and needs more review.