OpenAI integrates financial reasoning into main model training runs

OpenAI has resolved pre-training problems that previously hindered model scaling, with improvements noted in 2026. The company is integrating financial reasoning data upstream into main model training runs, such as GPT-6 Astra, ensuring this intelligence is accessible to third-party builders rather than limited to first-party products. Joseph Kim noted that existing financial benchmarks are saturated or broken, with incorrect answer keys and poorly specified prompts driving errors. Consequently, OpenAI stopped tracking the GDP eval for GPT-6 Astra because it was completely saturated. The company is now developing new, profession-specific evaluations, such as an investment banking eval, to better predict the ability to automate specific professional tasks. OpenAI expects models to improve significantly in human taste and aesthetic judgment by 2027, using human preference data to train reward models for softer objective functions. This approach aims to advance beyond basic factual accuracy by training models on human preference for aesthetic and structural quality.