THE CRUNCH
OpenAI's GPT-6 Astra has topped the ErdosBench for open math problems, a result that appears to contradict the company's stated priorities. Chief scientist Jakub Pachocki revealed that mathematics was deliberately not a priority for the model. Instead, OpenAI is directing resources toward recursive self-improvement and alignment research. This outcome supports the theory of a 'spiky' AI development path, where models
achieve extreme strength in select domains rather than broad progress. The performance suggests that targeted optimisation with human-generated data can still yield surprising results even when a specific capability is not the primary engineering goal.