Hypatia, Alverà's start-up beats AI bigwigs

A year ago, AI stocks, and the financial markets more generally, were shaken by the DeepSeek 'cyclone', the Chinese start-up apparently able to rival the most advanced American systems at a fraction of the development cost and computing power used by other companies. Today, with due proportions, there is a Milanese technology boutique that, in a 'competition' organised by enterprise software giant ServiceNow, has outperformed the flagship models of Google and OpenAI. We are talking about Ipazia, which was founded in 2021 by Giorgio Alverà, a former Goldman Sachs manager and brother of Marco (former CEO of Snam who is also a partner in Ipazia itself) and which - according to the press release - has outperformed the American AI bigwigs in WorkArena++, an industry benchmark for testing the ability of language models to solve real problems. It is not a matter of writing a poem or summarising a text, but of performing complex business processes: browsing product lists, ordering items, retrieving data from heterogeneous databases and filling out forms. In this scenario, called L1, - it is explained - Hypatia scored 90.3%, beating out Gemini-3 Flash from Google (86.1%), GPT-5 from OpenAI (79.1%) and Claude-4-Sonnet from Anthropic (63.3%).

"We have not changed the basic models, we have made them smarter in the way we use them," they explain from the company. The result "is a reasoning capability that makes advanced AI feasible for companies, with certain costs and time". In particular, they point out, Ipazia has developed a technology of'AI agents' that act like a team of experts: an added layer of intelligence that allows a complex problem to be broken down into manageable sub-tasks.

Behind Ipazia is a team of 18 professionals with seven PhDs and profiles from MIT, Microsoft, Goldman Sachs and Fondazione Bruno Kessler. In 2024, it was the only Italian startup among the 15 signatories of the G7 Declaration on AI, sitting at the same table as Amazon and Meta, and its technology is already operational in the banking sector, in recruiting, and in so-called 'responsible gaming', where it develops algorithms that identify risky behaviour to protect users.