OpenAI launched the public beta of its Agents API on September 10, opening it to all developers. The API packages the agentic runtime that powers Codex and ChatGPT — what OpenAI internally calls the Codex harness — into a hosted service: the loop of calling models, using tools, and managing context now runs on OpenAI's own infrastructure, while developers decide which capabilities an agent can use and where its code executes.
There is no added platform fee. Developers pay only for tokens, tool calls, and container runtime. Hosted sandboxes are billed at standard container rates, and model usage is billed separately at normal API prices; the announcement did not list a standalone price for the sandbox itself.
Four built-in capabilities
According to OpenAI, the API now handles several things developers previously had to build themselves:
- Automatic context compaction: as a session approaches the context limit, earlier content is compressed automatically;
- Tool search: tool definitions load only when needed, saving both tokens and cost;
- Programmatic tool calling: supports parallel and chained calls;
- Multi-agent orchestration: a primary agent can hand off complex sub-tasks to independent sub-agents.
There are three options for where an agent actually runs. The first is an OpenAI-hosted sandbox, which can be preloaded with files and dependencies and equipped with skills and plugins, letting an agent run code, process files, and produce finished output. The second is connecting over WebSocket to a developer's own machine. The third is routing to a partner sandbox — the current list includes Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel.
Numbers from early customers
OpenAI cited results from several early users. Ciridae's evaluation score rose from 0.71 to 0.85, and latency for its sub-agent workflow dropped to a quarter of what it was; SafetyKit's costs fell 60% after migrating; Hypha's failed responses dropped 86%; and Nash.ai is running several thousand long-running agents across a global logistics network.
These figures are self-reported by the customers and published by OpenAI, without independent verification. The announcement does not say what each company used before migrating or how the baselines were set, so the numbers should not be compared directly against one another.
Three generations of OpenAI's agent interface
Placed against OpenAI's own product history, the trajectory is fairly clear.
In late 2023, OpenAI launched the Assistants API, bundling conversation threads, file retrieval, and a code interpreter into server-side state that developers could call through an API — convenient, but opaque and hard to debug; a number of teams used it for a while and then rebuilt their own stack. In March 2025 it was replaced by the Responses API paired with the open-source Agents SDK, handing state management and orchestration logic back to developers to write into their own code. AgentKit followed that October, adding visual orchestration and evaluation tools on top.
The Agents API pulls back in the other direction: the loop moves back onto OpenAI's servers, while the execution environment stays open to third parties and to a developer's own hardware. What the first two generations required developers to build themselves — scheduling, compaction, sandboxing — is now the default.
The positioning suggests the framework behind Codex is no longer just for coding scenarios; OpenAI is packaging what it runs internally and offering it externally. What developers save on orchestration engineering, they trade for deeper dependence on OpenAI's runtime: how much logging is visible when the hosted loop fails, and whether a full trace can be exported, is not yet spelled out in the beta documentation.
OpenAI says it will iterate quickly on developer feedback during the beta before moving to general availability. When GA will arrive, and whether pricing will change at that point, has not been announced.
Sources: OpenAI's official announcement, OpenAI developer community, CocoLoop, MarkTechPost; pricing, the partner sandbox list, and early customer figures were checked against OpenAI's announcement.