Amgen puts GPT-5 to work on drug discovery data

Amgen is putting OpenAI’s new GPT-5 model through its paces, and the early word from the biotech giant is that it’s a significant step up for handling the messy, deeply technical data that drives drug discovery. The company, one of the first large enterprises to publicly detail its GPT-5 usage, isn’t just kicking the tires. They’re using it to wrangle complex biomedical information, moving well past the simple Q&A chatbots that defined the first wave of enterprise AI.

The core challenge Amgen faces isn’t generating text; it’s structuring and understanding data that is inherently unstructured. Think clinical trial reports, genomic databases, and decades of internal research documentation. According to the company, GPT-5 shows a markedly better ability to connect disparate data points and handle multi-step reasoning without falling apart. “What we’ve seen with GPT-5 is a more reliable ability to follow complex, domain-specific instructions,” an Amgen data science lead noted, pointing to the model’s improved context handling as a key differentiator from its predecessor.

This isn’t about replacing scientists. The workflow Amgen describes is more akin to a hyper-competent research assistant that can synthesize information across silos in seconds. A researcher can ask for a summary of all known drug interactions for a specific target across internal and external data, and GPT-5 can pull it together with fewer hallucinations and a clearer citation trail. The practical upshot is a potential acceleration of the hypothesis-generation phase, which is often a gating factor in early-stage R&D. Less time hunting for data means more time thinking about what the data means.

The financial stakes here are massive. Drug development is a famously expensive and failure-prone process. If a model can reliably highlight dead ends or surface non-obvious connections, the value proposition extends far beyond a productivity boost. Amgen’s willingness to go on the record signals a broader shift: the most demanding enterprise customers are now judging AI models not by clever chat replies, but by their ability to handle real scientific rigor without breaking.