27B model beats Opus and GPT at research

Inherent, a London startup founded by Google DeepMind alumni including chief scientist Edward Hughes, built Faraday, a 27-billion-parameter agent trained to independently reproduce the findings of published science papers, and it beat both Claude Opus 4.8 and GPT-5.5 at the task, TechCrunch reported August 22 citing Inherent’s own research writeup. Inherent trained Faraday with “Replica,” a reinforcement-learning environment of 310 tasks pulled from 100 papers spanning NLP, structural biology, materials science, and weather forecasting, using per-task rubrics and turn-level credit assignment to keep the reward signal stable over long-horizon runs. Faraday calls GPT-5.5 Codex as a coding tool during its own runs, and its advantage over the larger comparison models was most pronounced in meta-learning, structural biology, and materials science. Inherent raised a $50 million seed round when it emerged from stealth in May.

What it means for you

Faraday’s win margin comes from training against a narrow, well-specified reward signal, not from having more parameters, which is the same lesson behind routing tasks between AI models rather than defaulting to whichever frontier model has the biggest number attached. If you’re building or evaluating agents for a bounded, evaluable task, especially one where you can define a clean success rubric the way Inherent did for paper replication, a smaller specialist trained against that rubric is worth testing against a general frontier model before you pay frontier-model prices. It’s also a data point for why narrow, vertical agents keep beating general ones at the specific job they were built for.