C5R's AI-run lab scores 45% on its own benchmark
C5R built Facility-0 in 12 weeks, a lab where an AI model runs the full experimental chain from checking reagents to issuing technician instructions, and its SciUniverse Level 1 benchmark found Claude Fable 5.1 succeeded on 45.3% of 92 real scientific tasks on the first attempt, the best of six models tested.
C5R, a San Francisco startup, built Facility-0 in 12 weeks: a laboratory where the AI model itself picks the next experiment, controls some instruments directly, and issues written instructions to human technicians for everything else. On September 24 the company published SciUniverse Level 1, a benchmark of 92 tasks in 17 groups across chemistry, biology, and materials science, from compound synthesis and PCR (polymerase chain reaction) to pressing ceramic pellets. The top performer was Claude Fable 5.1 at 45.3% first-attempt success. All five other tested models scored lower.
The model receives a scientific goal (synthesize a compound, read a spectrum) and independently follows the full researcher workflow: it surveys available reagent inventory and instruments, designs the experiment as code, and that code is translated into commands for equipment and text instructions for the technicians. After the experiment the model reads the measurements and decides what to try next, accounting for remaining reagents and the previous outcome. Some tasks ran physically in Facility-0; others ran on a digital twin of the laboratory, a software model loaded with real experimental data.
Models handled experimental strategy and calculations competently. The failures clustered in physical, hands-on protocol details: they reused a single pipette tip across different DNA-containing wells, cross-contaminating samples; vortexed plates with uncovered wells; pipetted samples that had not yet thawed; and failed to account for solvent evaporation during prolonged reactions.
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 45.3% figure is a result on this specific test and task set; there is no basis for generalizing it to AI scientific work overall.
C5R founder 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: first running planarian experiments in his apartment, then joining a lab at the University of California, San Francisco (UCSF). "Models were making progress in the virtual world but fell short in the laboratory. We founded C5R to fix that." The UCSF page on Eternal Search: University of California, San Francisco.
By its competitors' standards C5R is small. Its only known institutional investor is early-stage fund Mythos Ventures, with the remainder 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 plus a separate DARPA contract for an autonomous laboratory, though at the time no independent benchmark for such claims existed.
C5R considers SciUniverse Level 1 a first iteration. Future versions will cover more complex tasks.
The item is concrete and dateable (September 24 publication), with a specific numeric result, a named failure mode, and a real institutional connection. The UCSF link is factual and traceable: founder Acilian joined a UCSF lab before starting C5R. The page gives readers an organizational entry point to that context without overstating the university's role in C5R's work.