This Day in AI Podcast We Committed Fraud with OpenAI's New Image Model (and Called Mum) - EP99.38
118 snips
Apr 24, 2026 They unpack recent big model drops and why some feel like vaporware. The conversation focuses on a terrifying image model that produced hyperreal forgeries. They debate token economics, hidden cloud costs, and how agents drive pricing way up. Short takes on underrated smaller models, the everything app race, and why automation in enterprises is harder than hype suggests.
AI Snips
Chapters
Transcript
Episode notes
OpenAI Pushing Models Into Apps To Build A Super App
- OpenAI is prioritizing putting new models into its consumer apps to build a super-app habit rather than immediately opening them via API.
- Michael notes GPT-5.5 launched to apps first and isn't available in the API, showing a product-layer strategy to lock users into interfaces.
AI Token Pricing Is A Heavily Subsidized Illusion
- Token economics are heavily subsidized: VCs and cloud credits cover most real costs while consumers pay a tiny fraction.
- Chris reports VCs pay ~70% of token cost, consumers pay ~5.5%, and enterprises bear most real cost.
Budget For Agentic Tasks Not Just Chat
- Expect agentic tasks to cost 10–50x more tokens than simple chat because of planning, tool calls and reasoning steps.
- Chris measured agent runs using thousands of input tokens and tens of thousands of output tokens for complex tasks.