OpenAI Starts Letting Enterprise Customers Pay Only When the AI Works

OpenAI has started letting some of its largest customers pay only when the AI actually completes the job, according to The Information’s report today. The arrangements cover select major accounts rather than the general public, with an end-to-end customer support interaction as the cited example. Nothing has been announced publicly, and OpenAI’s terms, customers, and prices are unknown. But the direction is unmistakable: the largest model provider is testing contracts where it absorbs the cost of failure.

Why Token Billing Is Losing the Argument

Token-based pricing charges for every attempt, including the failed ones, which puts all the model risk on the buyer. Outcome pricing flips that: the vendor only gets paid on completed results, so success rates become the vendor’s problem. The Next Web’s analysis of the report notes this model has been standard in customer support AI for about a year, and the benchmark numbers are already public. Intercom charges $0.99 per conversation its Fin agent resolves, and nothing when it does not. Zendesk has billed only for “Verified Resolutions” since May, roughly $1.20 to $1.50 each on committed volume, confirmed by an LLM evaluation within 72 hours. Salesforce went the other direction: Agentforce’s flat $2 per session drew complaints about cost unpredictability, and its later Flex Credits still bill failed actions.

The buyer preference data explains why OpenAI is moving. A Futurum Group survey cited in the analysis found 43% of AI buyers prefer consumption-based pricing and 27% prefer outcome-based, while fewer than one in five still want per-seat licenses. Unpredictable bills are a known enterprise-deal killer, and outcome pricing removes the objection that stalls pilots.

The Catch: Defining Done

The hard part is not the pricing, it is the definition. A resolved support ticket is measurable. Multi-step agentic work is not: OpenAI claims around 10 million agent users, and whether an agent “completed” a research, coding, or operations task is often a matter of judgment rather than a field in a database. Vendors also have to price failure in somewhere, which is why per-resolution prices cluster around a dollar rather than a cent, and why the model favors providers confident in their success rates. Token billing’s own horror stories push buyers toward the experiment anyway; one developer running 100 parallel agents famously accrued $1.3 million in tokens in 30 days.

For teams building on OpenAI, expect pricing conversations to shift from cost-per-token to cost-per-outcome, and expect contracts to hinge on how rigorously “success” is specified. The teams that win under this model will be the ones whose agent evaluations, the same benchmarks used to prove a task was done, are built to survive a negotiation. It rhymes with the shift already visible in agent task costs falling across the industry: as inference gets cheaper, the value moves from generating tokens to guaranteeing results.