OpenAI is falling short of its own projections as it races toward an IPO. The Wall Street Journal reports the company missed key revenue and user targets, an uncomfortable reality for a startup valued at tens of billions. The math isn't working.
The economics keep getting worse as AI evolves past simple chatbots. "Thinking models" like GPT-4 burn through tokens at a much higher rate. Now the push toward agents and tool-calling drives computational costs even higher. Long-horizon agent sessions that plan and execute multi-step tasks consume massive amounts of tokens. And no one has figured out how to make those economics work at scale.
Agentic swarms aren't ready. Current models can't reliably produce planning documents, create verifiable artifacts, break down tasks to minimize risk, or recognize their own limitations. AutoGPT proved the concept was possible but ran face-first into the same problems, and LangChain's multi-agent orchestration only gets you so far when the underlying agents struggle with basic task decomposition.
CognitionAI's Devin, a coding-specific agent, can autonomously complete development tasks within its narrow domain. Agentic coding tools like Steve Yegge's Gas Town are shipping to fill this gap, while specialized agents in vertical domains prove more viable than general-purpose swarms. Microsoft is baking agents into enterprise workflows through Power Platform, keeping expectations modest. General-purpose agent swarms remain expensive and unreliable. The gap between impressive AI demos and products people will pay for remains stubbornly wide. OpenAI's balance sheet is proof.