The models are already this powerful—why doesn’t OpenAI just start 100 companies? @OmniTools_AI summarizes a trending Hacker News thread: If frontier models are this strong, why is OpenAI still selling tokens instead of building applications and capturing the end-user profits? 📌 Key Takeaways 1️⃣ Models can already get real work done: OpenAI claims GPT-5.5 achieves or exceeds industry expert performance on 84.9% of 44 occupational tasks in GDPval; CFO Sarah Friar notes that Canva’s code is “100% generated by OpenAI,” and Travelers Insurance’s AI claims assistant scaled nationwide within two months. 2️⃣ What’s truly lacking isn’t intelligence, but “who is accountable for the outcome”: Models excel at well-defined tasks, but running a business requires market judgment, cash flow management, regulatory compliance, and handling customer complaints—all with financial consequences for mistakes. The community consensus is: no one dares to let a model be held accountable for results. 3️⃣ Codex shifted to strict token-based pricing: Moving from monthly estimates to usage-based API billing means OpenAI is selling access to compute and capability, not outcomes—the risk remains with the user. 4️⃣ This business is already huge: Friar revealed OpenAI’s annualized revenue run rate surpassed ~$40 billion as of July 2026, with enterprise customers accounting for over half. Ramp data shows enterprise spending on OpenAI APIs grew 82% quarter-over-quarter. 5️⃣ “Pay-for-performance” is still in the future: Friar says pricing based on business outcomes may come later, but Codex’s current model focuses on granular token billing. HubSpot has begun trialing “pay only if successful” models, though limited to easily measurable use cases like customer service and sales lead generation. In short: The models are powerful, but OpenAI has chosen a safer path—charging for capability while leaving operational risk with its customers.