OpenAI launches the Agents API and turns the Codex harness into a service

September 3, 2026. OpenAI launches GPT-6 Astra, its newest model. September 10, 2026. The company opens the Agents API in public beta and, the same day, pauses new sign-ups for its $200-a-month ChatGPT Pro plan for lack of capacity. September 6, 2026. An internal research report reveals that OpenAI’s researchers were already logging 3.1 agent-workdays for every human workday by mid-August. These three dates tell the same story: the barrier holding back agentic AI is no longer model quality — it is the cost and capacity of the inference running behind it.

The Codex harness, sold by the piece

The Agents API introduces no new model. It exposes the backend of Codex — the machinery that keeps an agent alive for hours — as a service. Until now, a developer who wanted to run an unattended agent had to build their own system to track the work. The API does it for them: it tracks the job as it progresses and gives the agent somewhere to execute its work, even when a task stretches well beyond a single context window.

The mechanisms are the ones Codex already used internally. As a task lengthens, the API can compact earlier context, so the agent does not stop dead at the model’s context limit. It can invoke tools only when needed, or hand parts of a larger job to subagents running in parallel. The actual work can run in OpenAI’s sandbox or on infrastructure the developer controls.

The commercial point is as sharp as the technical one: the orchestration layer leaves the bill. The developer pays for the models, tools, and hosted compute they actually consume — no longer the plumbing that keeps the agent standing. The promise is a collapsed cost of experimentation, and therefore an adoption curve that opens up.

The bill for agent inference

The tradeoff is that it becomes easier to consume more inference. As an agent progresses, it goes back to the model for the next step; a task that runs for hours can rack up far more inference than a typical API call. And when agents work in parallel, usage climbs even faster.

OpenAI has already lived this inside its own walls. The research report published on September 6 says the research organization was logging 3.1 agent-workdays per human workday by mid-August, measured in standard eight-hour equivalents. The median researcher, ranked by agent usage, was spending more than $600 a day on inference at API prices, while the 90th percentile exceeded $7,000 a day. Before June, researchers were still putting in more hours than their agents; by mid-August, agents were doing three times as much work.

OpenAI’s researchers are an extreme case, but the numbers show what happens when agent use starts to scale: one person can suddenly generate far more inference than their headcount would suggest. That is the real economic novelty of the moment, and it matters more than any benchmark.

The friction that was holding demand back

The Agents API lowers the cost of that experimentation by taking the orchestration out of the bill — but it also removes a friction that had been mechanically holding consumption back. Context compaction is the cleanest example. Before, a full context window forced the developer to decide what to discard or how to summarize the work so far. Now the API handles it automatically and the agent keeps going. That is useful, but it also means the workload no longer stops when the context fills up.

The shift is subtle and important: you are no longer selling a model call, you are selling a duration of autonomous execution. The more agents there are and the longer they run, the faster usage accumulates. One developer can have several agents at once, each returning to the model throughout the task — so headcount no longer tells you anything about compute consumption.

Astra hit the ceiling

The Astra rollout offers a preview of what that looks like. OpenAI stopped accepting new Pro subscribers less than two weeks after the model launched on September 3, saying those accounts put “the most strain on our systems.” Thibault Sottiaux, engineering lead for Codex, wrote on X that OpenAI was working to add capacity “as fast as we can.”

The Agents API and ChatGPT Pro are separate products, with their own rate limits and usage tiers: nothing indicates one is taking capacity from the other. Still, the calendar coincidence is telling. The company is making it easier to run agents for hours or days while pulling back access to its heaviest-use consumer plan, all while adding capacity. That is the sign the constraint has changed nature: the fight is no longer for a better model, it is for running the one you have.

Cloudflare made the same bet this summer, arguing that the infrastructure around AI workloads would eventually matter as much as the models themselves. OpenAI’s Pro pause is the concrete proof, barely two weeks after a model release.

What this changes for you

For an engineering lead or an SRE, the episode is a warning signal more than a product announcement.

• Budget inference as a duration, not a call volume. An agent running for hours consumes like hundreds of ordinary calls; the bill is thought in agent-days, not single tokens.
• Treat capacity as a product risk. When OpenAI freezes a plan for lack of capacity two weeks after a launch, availability becomes a design variable, not an ops detail.
• Keep control of the execution infrastructure. The API lets work run on your own infrastructure: that is the lever for controlling cost and latency when consumption takes off.
• Measure usage per person, not per seat. OpenAI’s report — 3.1 agent-workdays per human workday — shows a single engineer can multiply inference without the org chart moving.

The deeper trend outlasts the product. The economics of agents are shifting from “which model is smartest” to “who can run the most agents at the best cost.” Those who build capacity and accounting now will be the ones who can adopt without hitting a ceiling.

Verdict

The Agents API is a strategic move more than a technical launch: OpenAI is choosing to sell orchestration as nearly free in order to capture the inference consumption it releases. Its own researchers already show what that consumption looks like at scale — $600 a day in median, more than $7,000 at the 90th percentile.

If you build agents, the API is worth a trial: it removes the orchestration plumbing and makes long-running accessible, but keep your hand on the execution infrastructure and set cost guardrails from day one. If you operate the platform those agents will run on, read the Pro pause as a warning: inference capacity is becoming the new bottleneck, and it is paid day by day, not by benchmark.

References

• The New Stack — OpenAI’s researchers burned $7,000 a day on AI agents — now it’s opening the floodgates, September 11, 2026
• OpenAI — Introducing the Agents API, September 10, 2026
• OpenAI — research report on agent usage (3.1 agent-workdays per human workday), September 6, 2026
• Thibault Sottiaux (@thsottiaux) — announcement of the Pro plan pause, September 10, 2026