The Next Great Leap in AI Is Behind Schedule and Crazy Expensive

Abstract

OpenAI’s GPT-5 project, code-named Orion, faces significant technical and resource challenges that have delayed its development beyond mid-2024 expectations. Multiple training runs, each costing approximately half a billion dollars, have yielded improvements over current models but not enough to justify operational costs. Key obstacles include insufficient high-quality training data from public sources, leading OpenAI to generate custom data through human experts and AI systems. The company’s traditional strategy of using increasingly larger datasets shows diminishing returns, prompting exploration of alternative approaches like reasoning models. Internal turmoil, including leadership changes and staff departures, has further complicated development. The project’s struggles reflect broader industry concerns about potential plateaus in AI advancement and limitations of current training methodologies. Despite these challenges, OpenAI continues to pursue GPT-5’s development through combined approaches of reasoning models, AI-generated data, and human-created content. - AI-generated abstract.