OpenAI Tests 'Pay When It Works' Pricing as Agent Fails 62% of Tasks

GaryMarcus · x · 2026-09-01

OpenAI is testing a pricing model where enterprise clients only pay when AI agents successfully complete tasks, with OpenAI absorbing compute costs for failures. Gary Marcus suggests this shift is driven by customers refusing to pay for unreliable outputs. Cited data reveals OpenAI's Operator agent fails 62% of real-world desktop tasks.

Related event: OpenAI tests outcome-based pricing: pay only when AI succeeds(4 posts)→

Original post →

More from Companies & People

Companies & People channel →