Patronus AI

Patronus AI has changed what it sells. The site now presents the company as a research laboratory that develops training data, tasks and simulated environments for frontier AI laboratories, aimed at teams doing mid-training and reinforcement learning rather than at teams evaluating their own applications. The work is organised into three offerings — tasks, environments and research — with published benchmark papers covering world modelling, long-term memory in changing environments, multi-turn search with incomplete information, failure localisation in agent workflows and financial reasoning against real company documents.

That suits a small number of buyers: frontier laboratories that need expert-grounded tasks, graded rubrics and simulated applications to train agents on. It suits far less well the reader this entry used to describe — an application team wanting an evaluator for its own prompts and outputs — because the current pricing page does not publish a product to buy. It states that the company supplies training data to frontier laboratories and asks visitors to make contact to discuss scope and price, and the site carries a separate link to an archived pricing page, which is the shape of a retired product line rather than a current one. Guides on language-model testing and agent reliability remain online, which is evidence of what the company used to do rather than of what it now offers.

Not verified: which, if any, of the old evaluation product remains available could not be established. The application subdomain returned an HTTP 400 error and the documentation subdomain returned HTTP 401 when fetched, and no account was opened, so whether the old platform still runs is unknown. The benchmark results, the customer counts and the research claims are the company's own material on its own pages. No task, environment or evaluation was used, so nothing about output quality was tested.