C5R has set up a physical laboratory driven by models in twelve weeks, and publishes the score they obtain

Facility-0 gives models control over 40 real instruments; on the SciUniverse bench, Claude reaches 45.3% compared to 32.5% for GPT-6.

The San Franciscan start-up C5R claims to have built a research facility in twelve weeks that can be fully controlled by state-of-the-art models. Called Facility-0, it covers protein design, medicinal chemistry and solid materials, and puts more than forty instruments, from hydraulic presses to individual pipettes, under software control.

The target loop is that of a researcher: starting from a goal, the model explores the available inventory, reads the specifications of the devices, designs its experiment in the form of code, which is translated into equipment orders and instructions for human operators. It then analyzes the measurements obtained, replans and decides on the next test. The challenge is not the automation of a known protocol, but the ability to decide what to try.

C5R accompanies the announcement of the SciUniverse evaluation bench, the first level of which brings together 92 tasks carried out in the real installation, in chemistry, biology and materials science. The first published scores place Claude at 45.3% and GPT-6 at 32.5% — figures that above all show how far the models remain from an autonomous experimental practice, on tasks that a qualified researcher can carry out without difficulty.