Developers can now access OpenAI's Agents API in public beta, gaining direct access to the orchestration layer that powers Codex. The service eliminates the need for custom infrastructure to maintain agent operations over extended durations, as the API itself manages job state and provides execution environments that persist beyond single context windows.
The launch arrives on the same day OpenAI halted new subscriptions to its $200-per-month Pro plan, citing strain from GPT-6 Astra demand. Thibault Sottiaux, engineering lead for Codex, stated on X that Pro subscriptions "put the most strain on our systems," noting that OpenAI was expanding capacity "as fast as we can." While the Agents API and ChatGPT Pro operate as distinct offerings, the concurrent timing underscores the company's infrastructure pressures as it simultaneously broadens developer access to computationally intensive agent workloads.
Agent inference adds up fast
https://x.com/thsottiaux/status/2098113585683808624?ref_src=twsrc%5Etfw
The API handles context compression as tasks extend, preventing agents from halting when reaching model limits. Tools can be invoked selectively, and larger jobs can be distributed across parallel subagents. Execution occurs either within OpenAI's managed sandbox or on developer-controlled infrastructure.
The actual work can run in OpenAI's sandbox or on infrastructure the developer controls.
As agents advance through tasks, they repeatedly query the model for subsequent steps. Multi-hour operations accumulate substantially more inference than typical single API calls, with parallel agent execution accelerating consumption further.
OpenAI's internal experience illustrates the scale potential. According to a research report released September 6, the organization's research division achieved 3.1 agent-workdays for every human workday by mid-August, measured in standard eight-hour units. The median researcher by agent usage spent over $600 daily on inference costs at API pricing, while the 90th percentile reached $7,000 per day.
Prior to June, human researchers still exceeded agent output, but by mid-August, agents were completing three times the work volume. Though OpenAI's research teams represent an extreme scenario, the figures demonstrate the computational implications when agent adoption scales. Individual users can generate inference volumes far exceeding what traditional headcount metrics would predict.
One person can suddenly generate far more inference than their headcount would suggest.
Friction limited compute demand
The Agents API reduces experimentation costs by excluding orchestration from billing. Developers pay exclusively for model inference, tool usage, and hosted compute their agents consume.
However, this design simultaneously lowers barriers to higher inference consumption. Context compaction exemplifies this dynamic. Previously, full context windows forced developers to choose between discarding information or summarizing progress. The API now handles this automatically, allowing agents to continue operating. While beneficial for developers, this removes the natural checkpoint that previously occurred at context limits.
Astra demand hit the ceiling
The Astra rollout provides insight into potential capacity constraints. OpenAI suspended new Pro subscriber registrations within two weeks of the model's September 3 launch, citing the strain these accounts impose on infrastructure. The Agents API operates under separate rate limits and usage tiers, so the Pro pause does not directly impact developers using the API. Nevertheless, the company is already managing capacity allocation around its newest model.
Infrastructure outweighs benchmarks now
Increased agent deployment and longer execution windows accelerate usage growth. A single developer might operate multiple agents simultaneously, each querying the model repeatedly throughout task execution. Headcount alone provides limited insight into actual compute consumption.
For extended-duration agents, the core challenge involves sustaining progress while avoiding token waste and maintaining task continuity. Cloudflare advanced a comparable thesis this summer, contending that infrastructure supporting AI workloads would eventually rival model performance in importance.
For long-running agents, the challenge is keeping the work moving without wasting tokens or losing track of the task.
Source: The New Stack