OpenAI opened the Agents API in public beta on September 10, 2026, exposing the managed Codex harness—sessions, orchestration, context compaction—through an API. Hosted sandboxes are the convenience bet; partner and self-hosted environments remain options. There is no separate Agents API fee; tokens, tools, and containers still bill.
Introducing the Agents API · OpenAI
Quick Take
- September 10, 2026: OpenAI announces the Agents API in public beta.
- OpenAI hosts and maintains the harness; developers choose compute—OpenAI-hosted sandbox, own infrastructure, or partner sandboxes.
- Company pricing line: no additional Agents API fee; pay for tokens and tools (and hosted containers when used).
What shipped in the beta
OpenAI’s announcement describes the Agents API as bringing the same harness and infrastructure that powers Codex to developers through a flexible API. In plain terms, OpenAI operates the agent loop—coordinating model calls, tools, and context—while developers supply the tools, knowledge, and workflows that make an agent specific.
Alongside the API, OpenAI introduced an OpenAI-hosted sandbox that reuses sandboxing infrastructure familiar from Codex and ChatGPT. The company says it provisions and manages that environment so agents can run code, work with files, and produce artifacts, optionally configured with files, packages, skills, and plugins.
Environment choice is the product surface
Hosted sandboxes are the on-ramp, not the only door. OpenAI’s packaging and developer docs emphasize choosing an environment type: OpenAI-managed, self-hosted / VPC-style deployments, or first-class partner integrations. Public explainers summarizing the launch list partners such as Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel—treat partner enumerations as company ecosystem claims and verify against current docs before operators depend on a name.
That flexibility is the honest story for builders who cannot put code or files in OpenAI’s cloud. It is also a reminder that “Agents API” is not synonymous with “everything runs on OpenAI GPUs.”
Sessions, tools, and what beta does not prove
OpenAI positions long-running sessions, context compaction, tool search, and multi-agent delegation as harness features developers previously had to assemble themselves. Those are product claims about managed behavior. They are not, by themselves, independent measurements of reliability, latency, or total cost of ownership for a given workload.
Secondary articles have circulated customer-style ROI anecdotes. For AI Shift News, keep those out of the lead unless OpenAI’s primary post states them; even then, attribute. The Agents API is also distinct from the open-source Agents SDK: the API means OpenAI runs the harness, while SDK users still operate orchestration themselves.
Pricing without the fine-print trap
OpenAI’s line that there is no additional fee for using the Agents API during public beta is easy to misread as “free agents.” The same materials say developers pay for tokens and tools as on the pricing page, and hosted container usage applies when you use OpenAI’s sandbox. Budget models still matter.
Bottom Line
The Agents API is OpenAI productizing the Codex harness and betting that hosted sandboxes plus partner environments remove a slab of agent infrastructure work. Cover the beta as packaging and access. Do not confuse a public beta with production proof, and do not invent a separate flat Agents API tax that OpenAI says is not there.