FrontierSeptember 10, 2026via The Decoder
GPT-6 Astra gives mathematicians a breather, and OpenAI says that's by design
Why it matters
OpenAI is explicitly choosing 'spiky' capability development—excelling in select domains while deprioritizing others—and shifting resources toward self-improvement and alignment. This signals a strategic pivot in how frontier labs are allocating training effort and research focus.
Key signals
- GPT-6 Astra tops ErdosBench for open math problems
- Math was deliberately not a priority for this model, per chief scientist Jakub Pachocki
- OpenAI prioritizing recursive self-improvement and alignment research over broad capability
- 'Spiky' AI development path emerging—extreme strength in select domains rather than uniform progress
- Strategy reflects constraint: AI still needs targeted human-generated data and cannot yet fully improve itself autonomously
The hook
GPT-6 Astra tops ErdosBench despite math being deliberately deprioritized—OpenAI's bet on recursive self-improvement over broad capability.
OpenAI's GPT-6 Astra tops the ErdosBench for open math problems, even though chief scientist Jakub Pachocki says math was deliberately not a priority. Instead, OpenAI is pouring resources into recursive self-improvement and alignment research. That supports the theory of an increasingly "spiky" AI d…