DeepMind Alumni's 27B-Parameter Faraday AI Outperforms Claude and GPT-5.5 at Scientific Research Replication

My take: A 27-billion-parameter model built specifically for science outperformed Claude Opus 4.8 and GPT-5.5 across every category of Replica, a new benchmark that tests whether an agent can reproduce a paper's experiments without seeing the original figures. Faraday, from Inherent (a London startup founded by ex-DeepMind researchers), showed a particularly pronounced advantage in structural biology and materials science. That said, Inherent designed both the model and the benchmark, which is reason enough to wait for independent validation before treating these numbers as definitive.

What this illustrates beyond the concrete data: general-purpose models are powerful, but when work requires deep reasoning in a highly specific technical domain, a model trained for that task can deliver real advantages. Model size matters less than specificity of training.

For any team using AI in technical research or specialized analysis, the practical lesson is that the catalog of options extends well beyond the most widely publicized models.

Do you know what specialized models exist in your field, or are you evaluating only the most widely marketed options?