OpenAI's Best Life Science AI Model Fails 63.9% of Benchmark, Highlighting Limits of Current Systems

This week, OpenAI released a new benchmark called LifeSciBench, a 750-task test designed to measure whether AI systems can support realistic life science research. The top-performing model, GPT-Rosalind, achieved a pass rate of just 36.1 percent, failing nearly two-thirds of the tasks.

The test goes beyond simple biology questions. It requires models to work with supporting documents, figures, and complex datasets. When tasks involved artifacts or URLs, GPT-Rosalind's pass rate dropped from 45.1 percent on text-only tasks to 28.1 percent.

Performance Plummets with Complex Inputs

The benchmark revealed a familiar weakness: AI systems perform better when everything is presented as text. Once forced to handle external data, performance falls off. OpenAI notes that models are becoming capable of scientific communication and evidence synthesis, but they remain far from autonomous scientists.

• GPT-Rosalind led all tested models but still failed 63.9 percent of benchmark tasks.
• Text-only tasks: 45.1 percent pass rate.
• Tasks requiring artifacts or URLs: 28.1 percent pass rate.
• LifeSciBench includes 750 tasks measuring realistic research support, not just fact recall.

Industry Warnings on AI Reliability

The findings arrive amid broader unease about AI's reliability. Harvard Business Review recently warned that AI-generated low-quality output, dubbed "workslop," is degrading the information companies rely on for decisions. A Microsoft executive acknowledged that humans are struggling to keep up with AI advancement, suggesting a narrowing window to understand AI before it becomes too complex to control.

Yann LeCun, former chief AI scientist at Meta, criticized the industry's trajectory, calling Elon Musk's xAI "kind of a failure" and suggesting that the entire AI sector may need a reset. The criticism points to a common theme: current AI models, including the most advanced, still suffer from fundamental limitations in reasoning and handling real-world complexity.

OpenAI's LifeSciBench is not intended to prove AI useless in research. The company argues that models can already assist scientists overwhelmed by information. But the benchmark serves as a reality check. Autonomous scientific discovery remains out of reach, and the gap between promise and performance is still wide.

What comes next is uncertain. Startups like 2Brains Inc., co-founded by tech pundit Robert Cringely, are attempting to solve LLM hallucinations through patented architectures. But until benchmarks like LifeSciBench show significant improvement, the industry's confidence in AI as a reliable research tool will remain limited.