The news
The OpenAI Agents API entered public beta on September 10, 2026, giving developers a managed version of the harness behind OpenAI's Codex coding agent. A harness is the software around a model that keeps a task running, calls tools, manages memory and recovers from failures.
According to OpenAI's developer documentation, the company now handles session orchestration, context compaction and recovery, while the customer's application provides the tools and chooses where the agent runs. Sessions persist across turns, so an agent can resume long work without rebuilding its context, and developers can stream progress and connect custom tools or Model Context Protocol (MCP) servers, a common standard for linking AI models to outside software and data.
Agents can run in OpenAI-hosted sandboxes, which are isolated computing environments, or on a customer's own infrastructure. Investing.com and MarkTechPost reported that nine partners have integrations to supply compute environments: Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop and Vercel. The harness is based on an open-source version of Codex that developers can review.
OpenAI charges no fee for the API itself. Its documentation says model usage is billed at standard API rates, with OpenAI tools and OpenAI-hosted sandboxes billed separately at their usual rates. Investing.com reported that OpenAI plans to refine the product with developer feedback before general availability.
The documentation also sets out a significant limit: the API currently supports data residency only in the United States and does not support Zero Data Retention, an arrangement under which a provider does not store customer prompts and outputs. OpenAI notes that choosing a self-hosted sandbox does not make the API eligible for Zero Data Retention.
One early user reported gains. Jack Weissenberger, chief technology officer of Ciridae, said the API raised his company's evaluation score to 0.85 from 0.71 and cut latency fourfold for workflows that use subagents; OpenAI published these customer-reported figures with the launch, and they are not independent benchmarks. On the same day, OpenAI made its GPT-Live 1 voice model generally available at $0.05 per minute, with backend model and tool use billed separately.
The numbers
- Platform fee for the Agents API
- None (model, tool and sandbox usage billed)
- Compute-environment partners at launch, per Investing.com and MarkTechPost
- 9
- Supported data residency
- United States only
- Ciridae evaluation score before and after (customer-reported, published by OpenAI)
- 0.71 to 0.85
- GPT-Live 1 voice price
- $0.05 per minute
Why CEOs should care
For CTOs and engineering leaders, the API takes over the hardest plumbing in agent projects: keeping long tasks alive, trimming context and recovering from errors. That can shorten build time, but it also moves core agent logic onto OpenAI's platform. Before committing, ask how sessions and saved state can be exported, what changes when the product leaves beta, and whether the open-source harness lets you run the same agents elsewhere if you switch model providers.
For CISOs and compliance teams, the data terms matter more than the features. With US-only residency and no Zero Data Retention, the beta is a poor fit for European personal data, regulated health or financial records, or any contract that bars a provider from storing prompts. Sessions persist by design, so set deletion policies and limit which systems and credentials an agent's sandbox can reach, whether it runs on OpenAI's servers or yours.
CFOs should note that no platform fee does not mean low cost. Long-running agents consume model tokens, tool calls and sandbox time continuously, and costs scale with how long and how often they run. Require per-agent budgets and usage monitoring before scaling pilots, and compare the bill against running the open-source harness on your own infrastructure.
The bigger picture
The launch shows the major model developers competing to own the layer where business agents actually run, not just the models inside them. By packaging the harness behind Codex as a product, pairing it with cloud and sandbox partners, and releasing a voice model the same day that hands off reasoning to backend agents, OpenAI is building a full stack for agents that live inside its platform. Buyers get speed; they also take on deeper dependence on a single vendor's runtime.
What’s next
Watch for the terms OpenAI sets at general availability, especially whether it adds non-US data residency and Zero Data Retention, and whether any platform fee appears. Rival developers' responses, and whether large enterprises move agent pilots onto the API, will show how much of the agent runtime market OpenAI can claim.
What “Fact-checked” means
Fact-checking means testing a story’s facts against the evidence before it is published. This story went through at least two separate checks before this version was published.
- What we checked
- Its names, figures, dates, job titles, quotes and who said what were checked against the story’s sources, including its main source where it could be opened. The headline was checked for accuracy and overstatement.
- How
- A first check reviewed the whole story. If it passed, a second, skeptical check went back to the sources to look for mistakes in the most important facts. If a check flagged the story, it was edited to fix the problems found, and a separate re-check then reviewed the whole story again.
- Who
- The checks are made with our newsroom’s technology tools, as steps kept separate from the writing, under rules set by our editor, Hussein Mukhtar. A story the checks still flag is not published automatically; it is held for the editor, who decides whether it is fixed, published or dropped.
- If something is wrong
- “Fact-checked” does not mean error-free. If a material error is found after publication, we correct the story and add a note saying what changed. Report an error
Companies in this story
