OpenAI has opened its Agents API, handing outside developers the same foundation that Codex and ChatGPT run on. Agents can execute in sandboxes hosted by OpenAI or in ones provided by partners — Cloudflare, Vercel and Oracle. There is no separate charge for that infrastructure: billing depends only on tokens consumed. The API is built on the open Codex code scaffold and supports MCP, custom functions and built-in tools, web search among them.
Source: the-decoder.com
The pricing is the part worth stopping on. A sandbox is a machine that runs, and running machines are not free to operate. Folding that cost into the token price means OpenAI is choosing to be paid for model calls rather than for compute — which is the pricing a company adopts when it wants the execution layer to be a non-decision. Developers evaluating agent infrastructure now have one fewer line item to compare, and one fewer reason to build the sandbox themselves.
Giving a choice of partner sandboxes points the same way. An edge network, a deployment platform for web developers and an enterprise infrastructure vendor do not overlap much; what they share is that each already sits where some category of developer deploys code. Meeting all three is not a technical decision so much as a distribution one.
The claim doing the most work in the announcement is that this is the foundation Codex and ChatGPT are built on. Parity with a vendor's own products is the most persuasive thing an agent platform can offer and the hardest thing for a customer to verify. Building on the open Codex scaffold makes it checkable in a way that a marketing line would not, which is the strongest part of the release — the code that defines the agent loop is inspectable even if the model behind it is not.
What the announcement covers is choice and price. It does not address what happens when an agent running in someone else's sandbox misbehaves: what it can reach, how it is isolated, whose logs record it, who answers when an agent with a web-search tool and custom functions does something a user did not intend. Those are the questions that decide whether agent infrastructure is usable in production, and they are the ones a pricing announcement is structurally unable to answer.
Token-only billing also sets the incentives. An agent that retries, re-reads its own context and calls a tool five times is more expensive than one that gets it right the first time, and that expense is revenue. OpenAI has made its margin on agents a function of how much they think — and it is now selling the loop, not just the model that runs inside it.