OpenAI turns agent work into metered compute

OpenAI is turning agent work into a metered compute business, not just a chatbot business. The key shift is that developers can now launch jobs that keep running after a person closes the app, whether that means a coding task on an OpenAI managed computer, a research workflow that checks back later, or multiple agents working in parallel. That changes token demand from short bursts during active use into a steady stream of background usage tied to completed work.1 2 3 4 5

• The product plumbing now matches that revenue model. OpenAI documents background mode in the Responses API for long running tasks, and its Agents materials position the API as the main interface for agent style applications with tool use and orchestration built in.2 3 4
• Codex shows what this looks like in practice. Codex Cloud runs tasks on OpenAI managed computers, and by June 2026 the 99th percentile of Codex users were generating more than 60 hours of agent work per day across parallel agents, far beyond one person’s keyboard time.1 5 6
• This also mirrors Anthropic’s move from a model API into delegated work products. Anthropic positioned Claude Cowork as a way to hand off projects to Claude across Claude.ai, Claude Code, and the Claude Platform, showing that both labs are pushing from single prompts toward always on task execution.7 8

The next step is that API competition shifts from who has the best model to who makes autonomous work easiest to deploy, monitor, and scale. If OpenAI keeps collapsing the setup needed for background jobs, tool calling, and cloud execution, more software categories start to look like token pipelines that run all day instead of seat based apps people open for a few minutes.2 3 4 5