Frontier AI labs price subsidized user access below cost, betting on data from millions of free and cheap paid users to train the next generation of models before human feedback becomes unnecessary.

Current pricing—$20/month for individuals, $30/month for pro tier—costs labs several multiples more in GPU-hours, electricity, and model wear per heavy user. OpenAI has admitted its $200/month enterprise plan loses money on the most active users; Anthropic's economics are no better; Google subsidizes Gemini against ad revenue; Meta gives Llama away outright.

Every user interaction—edits, regenerations, thumbs-up/down votes, question phrasing—produces training signal that the model cannot synthesize for itself. Reinforcement learning from human feedback (RLHF), using corrections from paying and free-tier users, is what separates a great model from a good one in 2026.

Synthetic data has crossed a quality threshold where models can generate, filter, and grade their own training data competitively with human input. Agentic self-play removes humans from evaluation loops in domains with clean ground truth (code, math, structured tool use). A reasonable estimate places the inflection—where human-in-the-loop signal stops being necessary—inside three to five years for general-purpose assistants, faster for verticals with measurable success criteria.

After that inflection, the subsidy that currently justifies free and cheap access ceases to be justified. Consumer pricing is predicted to rise to $80–$150 per month (reflecting actual cost), most current users will churn, the strongest capabilities will gate behind five-figure-minimum enterprise contracts, and labs may vertically integrate to capture operator margins rather than license tools at any price.