OpenAI launched its Agents API in public beta on Thursday, letting developers run agents unattended for days while the company simultaneously halted new sign-ups for its $200-per-month Pro plan after demand for GPT-6 Astra overwhelmed capacity. According to a report published by The New Stack on September 10, 2026, the timing highlights a tension: OpenAI is making it simpler for developers to execute agents for extended periods while restricting access to its most resource-intensive consumer tier and racing to expand infrastructure. The report draws on OpenAI's own internal research showing how quickly agent usage can outpace traditional compute demands.
By mid-August, OpenAI's research organization was logging 3.1 agent-workdays for every human workday, measured in standard eight-hour equivalents, according to an OpenAI research report released September 6. The median researcher, ranked by agent usage, was burning through more than $600 daily on inference at API prices, while those at the 90th percentile exceeded $7,000. Before June, OpenAI's researchers were still putting in more hours than their agents, but by mid-August, the agents were performing three times as much work. OpenAI stopped accepting new Pro subscribers less than two weeks after GPT-6 Astra launched on September 3, with engineering lead for Codex Thibault Sottiaux writing on X that Pro subscriptions "put the most strain on our systems."
The report notes that the Agents API handles orchestration automatically, tracking jobs as they advance and providing agents a place to execute work even when tasks stretch far beyond a single context window. The API can compress earlier context so agents don't halt when hitting the model's context limit, bring in tools only when required, or dispatch portions of a larger job to subagents operating in parallel. Developers pay only for the models, tools, and hosted compute their agents actually consume, with the orchestration layer excluded from the bill. The actual work can execute in OpenAI's sandbox or on infrastructure the developer controls.
The report argues that as agents operate for longer durations, they cycle back to the model for the next step, and a job lasting hours can accumulate far more inference than a typical API call. The usage climbs even faster when agents work in parallel, meaning one person can suddenly generate far more inference than their headcount would suggest. Context compaction previously forced developers to decide what to discard or how to summarize work so far, but now the API handles that automatically and the agent continues, which is useful but also means the workload doesn't stop when the context window fills. The report concludes that for long-running agents, the challenge is keeping work moving without wasting tokens or losing track of the task, echoing Cloudflare's summer argument that infrastructure around AI workloads would eventually matter as much as the models themselves. OpenAI's own researchers represent an extreme case, but the numbers reveal what happens when agent use starts to scale—and headcount alone no longer tells you much about compute consumption. The real test will be whether enterprises can manage costs as orchestration friction drops and agents run longer, especially if capacity constraints persist across both developer and consumer tiers.