Dean W. Ball: Frontier-Model Economics Are 'a Bad State of Affairs' as Training Costs Get Recouped in a Shrinking Post-Release Window

Simon Willison / Dean W. Ball·medium signal

In a quote highlighted by Simon Willison, Dean W. Ball argues frontier models cost an enormous amount to train while only a fraction of that is recouped in the brief post-release window before the next model or a cheaper competitor erodes pricing power — which he frames as a structurally bad dynamic for the industry. The argument matters as a lens on the GPT-5.6 tiering and on who controls access to frontier capability, a question now spilling into policy debate. It's a sharper framing of the AI-economics narrative the audience tracks than the usual capex headlines.