ArcaScience launches Flow, claims its specialised AI beats frontier LLMs on drug safety
Paris- and Boston-based ArcaScience has launched Flow, an AI platform that helps pharmaceutical teams assess the benefit-risk profile of drug candidates. The company is also publishing a benchmark claiming its purpose-built system finds far more adverse-reaction evidence than general-purpose frontier models, at a fraction of the cost.
Startup Reporter has followed ArcaScience since its seed round last year. This launch is its biggest product step since then, and it tests an idea many in AI are watching closely: in regulated, high-stakes fields, a narrow system built for one job may beat the biggest general models.
Why benefit-risk assessment matters
Bringing a drug to market takes more than ten years. About 90% of candidates fail, and the average cost per approved drug is around $2.3 billion. Clément made the case for catching these problems earlier in an op-ed for Startup Reporter. Regulators such as the FDA and EMA require companies to show that a new drug’s benefits outweigh its risks compared with existing treatments.
The evidence for that case is spread across trials, publications and safety reports, so it is hard to compare scenarios or see where a candidate is weak. ArcaScience argues that many failures come from poor positioning: a promising molecule aimed at the wrong population, or safety issues found too late.
What ArcaScience Flow does
Flow brings together 24 specialised AI models and an evidence base the company puts at 100 billion biomedical data points. Clinical and safety teams can test different benefit-risk configurations, spot gaps in their evidence and decide what to investigate next. The analysis updates as new trial results come in.
The platform replaces ArcaScience’s earlier one-off query reports with interactive software. It can also draft regulatory documents, including PSUR/PBRER, Risk Management Plans and the benefit-risk section of eCTD Module 2.5. Every claim is traceable to its source passage, and the company says identical queries always return identical, auditable answers. That reproducibility matters for regulators and is a known weakness of general-purpose LLMs.
“A promising molecule deserves the strongest possible development strategy,” said CEO Romain Clément.
His motivation is personal, as he explained in our interview earlier this year: the company grew out of his own experience as a cancer patient.
The benchmark: specialised AI vs frontier LLMs
The company’s BRB-C v3.7 benchmark compared its core engine with five general-purpose frontier LLM APIs. ArcaScience reports these results:
- Recall: Flow’s engine found 104 of 108 adverse-reaction concepts (96.3%). The best-performing general model, Gemini 3.6 Flash, found 70 (64.8%).
- Cost: For a scenario covering 30 million articles, ArcaScience estimates its cost at $0.0162 per article, or $486K to $1.78M in total. It puts the cost of running general LLM APIs over the same dataset at $1.35M to $18.53M, with Claude Opus 5 at the top of that range.
The study was designed and run by ArcaScience. Consultancy inExtenso reviewed it qualitatively. The full methodology is published on ArcaScience’s website for readers who want to check the details.
Traction and funding
ArcaScience says it now works with more than 20 pharma clients, up from 10 when it raised its $7M (about €6M) seed round a year ago. Clients include Sanofi, AstraZeneca, GSK, Takeda and ICON, as well as the Paris Brain Institute (ICM). During the COVID-19 pandemic, the French government selected the company to structure the scientific literature on the virus.
The September 2025 round was led by The Moon Venture, with Pléiade Venture, Plug and Play Ventures, Bpifrance and AKKA Technologies taking part. Flow is available now for pharma companies, biotechs and regulatory teams.
Why it matters: Most AI coverage focuses on bigger general models. ArcaScience is betting on the opposite approach. If its numbers hold up, being deterministic, traceable and cheap at scale may matter more to pharma buyers than raw model size.
More on ArcaScience on Startup Reporter
0 comments