OpenAI launches Agents API public beta for cloud agents
On September 10, 2026, OpenAI announced the public beta of the Agents API. The interface is intended to provide developers with a managed layer for longer-running cloud agents, including tools, sandboxes, and subagent coordination.
On September 10, 2026, OpenAI announced the public beta of the Agents API. The interface is intended to let developers create and operate cloud agents with OpenAI’s managed Codex harness without having to build their own orchestration, context-management, and sandbox layers.
The Agents API is designed for workflows in which an agent does more than produce a single isolated model response: it works for longer periods, uses tools, and can coordinate multiple helper agents. Through the API, OpenAI is making available the core orchestration layer used in Codex.
What the OpenAI Agents API provides
According to OpenAI, the API supports long sessions with automatic context compaction. This feature is intended to manage context during longer-running tasks, where the volume of previous communication and working data would otherwise gradually increase.
The interface includes access to tools. OpenAI says it supports web search, MCP, and parallel programmatic tool calls. MCP, or Model Context Protocol, is used to connect models with external tools and data sources. The API also supports subagent coordination, meaning helper agents that perform parts of a larger task.
Developers can choose the environment in which the agent operates. Options include an OpenAI-hosted sandbox, their own infrastructure, or integrations with partner providers. The choice of environment is particularly important when working with internal systems, data, or custom operating rules.
Billing and operation in the public beta
OpenAI says it does not charge a separate fee for the Agents API itself. However, the tokens and tools used remain billable. For long sessions, the main factors affecting total costs will therefore be the models used, the extent of context processing, and calls to individual tools.
Because the service is in public beta, the interface, operational characteristics, and pricing terms for supported tools may change.
Why the Agents API matters to developers
Deploying agent workflows typically requires separately handling step orchestration, context management, tool execution, an isolated environment, and monitoring task runs. The OpenAI Agents API offers these components as a managed service, which could lower the technical barrier to building such systems.
Managed infrastructure does not replace organizational oversight, however. When deploying agents, organizations still need to define tool permissions, limit access to data, monitor costs, and verify agent outputs before using them in sensitive processes.
In its materials, OpenAI also cites customer claims of lower costs, latency, or failure rates. These are marketing references presented by the company, not independently verified comparisons. Independent tests have not yet confirmed the real-world reliability of the Agents API in production deployments or its impact on the agent-platform market.
What to watch next
- the general-availability timeline and possible interface changes,
- model, tool, and sandbox-operation pricing for long sessions,
- independent production tests of security, reliability, and costs,
- the scope and terms of integrations with external sandbox providers.
Sources
- OpenAI — Introducing the Agents API – Confirms the September 10, 2026 announcement, public beta, managed Codex harness, supported environments, and billing for tokens and tools.
- OpenAI API Docs — Agents API overview – Confirms the existence of product documentation and the scope of Agents API features, including sessions, sandboxes, tools, MCP, and multi-agent support.
Verified and updated: 09/11/2026 06:22
