Multiverse Computing launches Luminary to test language models across whole conversations

The story so far, in brief; the legwork, and the full article, stay with the original outlet.

Multiverse Computing, the San Sebastián AI company, launched Luminary on 21 September, a platform meant to tell businesses which large language model suits their particular job. It is open now through early access. The pitch starts from a selection problem: roughly 2.9 million models sit on public repositories, and the usual ways of ranking them rely on academic tasks and the price per token, neither of which is quite what a company is buying.

Luminary tests models on entire conversations instead of single prompts. Each candidate runs inside a digital twin that plays the customer over many turns and imitates the backend systems and tool calls, so lost context or drift from policy shows up before launch. Results are scored on cost per completed task, success rate, latency and policy compliance, with every test reproducible and an exportable audit trail. It covers Multiverse's own models and those of any OpenAI-compliant provider, and one possible answer is that a smaller, cheaper model will do.

"Teams pick the model with the loudest launch and discover the cost six months later," said Enrique Lizaso, co-founder and chief executive. Multiverse says it has more than 100 customers, among them Iberdrola, Bosch and the Bank of Canada.