OpenAI has started letting some customers pay only when the AI works

This is a pricing-power signal more than a revenue event. Moving from per-token economics to per-success economics tells you management believes enterprise buyers are already measuring AI against labor replacement, not software usage, which is usually the point where a model provider can start capturing a larger share of the value created. Near term, that can actually make reported revenue choppier because billing is contingent on task completion, but over time it supports higher enterprise willingness to deploy and higher lifetime value per account.

The clearest second-order winners are infrastructure and workflow owners that sit upstream of task completion: Microsoft/Azure, AWS, and GPU supply-chain names should see more inference intensity if customers scale end-to-end automation. The clearest losers are seat- or ticket-based vendors in customer support and business-process outsourcing, where buyers can now benchmark software directly against labor savings; that is a margin pressure vector for contact-center software and offshore service providers if this pricing model proves repeatable. If success-based pricing becomes a norm, it also compresses the moat for commoditized AI offerings, because customers will demand proof of economic output rather than model-quality rhetoric.

The main risk is measurement failure: if completion definitions are messy, collections get delayed, disputes rise, and OpenAI ends up warehousing downside on the hardest workflows. The key catalyst path is 1-3 months of pilot expansion and renewal language; the structural read-through is 6-18 months, when this either becomes the standard enterprise contract or gets rolled back to a niche offer. Consensus may be too focused on the bullish headline and not enough on the fact that the company is implicitly admitting token pricing under-monetizes high-ROI use cases; that is bullish for adoption, but not automatically bullish for near-term top-line visibility.