Startup C5R built an AI-run laboratory where the model directs both instruments and human technicians, and its SciUniverse test shows the best model fails more than half of real scientific tasks
Startup C5R built an AI-run laboratory where the model directs both instruments and human technicians, and its SciUniverse test shows the best model fails more than half of real scientific tasks
San Francisco startup C5R built Facility-0 in 12 weeks, a laboratory where the AI itself selects the next experiment, controls some instruments directly, and issues instructions to human technicians for everything else. On September 24 the company published the SciUniverse Level 1 test: 92 tasks spanning chemistry, biology, and materials science, from compound synthesis to working under reagent shortages. The top-performing model was Claude Fable 5.1 with a 45.3% first-attempt success rate; all five other models tested scored lower.
The model receives a scientific goal (for example, synthesize a compound or read a spectrum) and independently follows the full researcher workflow: it surveys the available inventory and instruments, designs the experiment as code, and that code is translated into commands for equipment and text instructions for the people who handle whatever is not yet directly automated. After the experiment the model reads the measurements and decides what to try next, accounting for the limited reagent supply and the outcome of the previous attempt. A single model call controls instruments, technicians, and the reagent store. Both the company and an independent review of the system confirm this.
The company explains why a test of this kind is needed: previous AI evaluations in science began after the experiment, on ready-made data, while the process of obtaining that data (choosing materials, operating instruments, responding to failures) was never measured. SciUniverse fills that gap.
The test evaluates the entire chain from goal to result across 92 tasks in 17 groups, from organic synthesis and DNA amplification by polymerase chain reaction to pressing ceramic pellets. Some tasks the model performs physically in Facility-0; others run on a digital twin of the laboratory loaded with real experimental data. Models handle experimental strategy and calculations competently but fail most often on physical, genuinely hands-on protocol details: they pipetted still-frozen samples, vortexed plates with uncovered wells, reused a single pipette tip across different DNA-containing wells (cross-contaminating samples), and failed to account for solvent evaporation during prolonged reactions.
Michael Acilian previously worked on wearable electronics at Apple and Misfit Wearables, then founded and sold Clara Labs, an AI meeting-scheduling service. A year and a half ago he moved into experimental biology: he set up a laboratory in his apartment for experiments on planarians, then joined a lab at the University of California, San Francisco:
"Models were making progress in the virtual world but fell short in the laboratory. We founded C5R to fix that."
By the standards of its competitors the company is small: C5R's only known institutional investor is the early-stage venture fund Mythos Ventures, with the remainder being roughly $900,000 from 44 individual investors. For comparison, Lila Sciences launched with $200 million in seed capital to build a scientific "superintelligence," and Medra raised $52 million in investment and separately received a contract from the defense agency DARPA for an autonomous laboratory, though at the time no independent benchmark for such claims existed.
The company has already named its next step: future versions of the test will address more complex tasks. C5R considers SciUniverse Level 1 only the first iteration of this kind of measurement.