What OpenAI released
OpenAI put the Agents API into public beta for all developers on 10 September, according to its API changelog. The service exposes the Codex harness, the software that runs OpenAI’s coding agent, as an OpenAI-managed API: developers supply the tools and the execution environment, and OpenAI runs the agent loop around them.
The split is spelled out in the developer documentation. OpenAI handles sessions, orchestration, context compaction and recovery, the plumbing that most teams building agents have so far written and maintained themselves. The harness is open source, KuCoin reported, so developers can read the logic that coordinates model calls, tools and context rather than treat it as a black box.
Sessions, sandboxes and subagents
The core unit is a durable session. An agent instance keeps its context across turns, so an application does not rebuild the conversation on every call, and a session can resume from where it stopped. When a long task outgrows the model’s working context, the harness compacts it.
Agents can work inside sandboxes, where they run code, edit files, connect to MCP servers and produce artifacts. They can load skills from capability directories and delegate pieces of work to subagents. OpenAI offers hosted sandboxes, and developers can also bring their own: KuCoin listed sandbox integrations with Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop and Vercel, and BigGo Finance also named Oracle and Cloudflare among the self-hosted options.
The code samples in the documentation use GPT-6 Astra, the model OpenAI released on 3 September and describes as its most capable.
What it costs, and where the data lives
There is no separate fee for the API itself. The documentation says model usage is billed at the selected model’s API rates, OpenAI tools at their standard rates and OpenAI-hosted sandboxes at standard container rates.
The limits matter more to enterprise buyers than the price. The Agents API currently supports data residency only in the United States and does not support Zero Data Retention, and the documentation states that choosing a self-hosted sandbox does not make a workload eligible for it. For European companies with residency obligations, and for anyone whose contracts depend on ZDR, that rules out production use for now, wherever the code executes.
GPT-Live-1 arrives in the API
The same day, OpenAI made GPT-Live-1 generally available in the API. It is a full-duplex voice model: it listens while it speaks, handles interruptions, and can hand reasoning and tool use to a backend model or agent while the conversation continues. The model first shipped inside ChatGPT Voice on 8 July, replacing the turn-based Advanced Voice Mode.
According to the model page, GPT-Live-1 costs $0.05 per minute, billed per second without rounding up to whole minutes; the backend model and tools are billed separately. It takes audio and text in and out, supports function calling and runs only on a new Live sessions endpoint. TestingCatalog reported 12 voices at launch, with custom voices available through OpenAI’s sales team.
The performance claims come from the launch, not from independent testing. Speak, the language-learning app, reported almost 80% fewer interruptions than with previous turn-based systems, and the model scored 30 percentage points higher than GPT-Realtime-2.1 on Full Duplex Bench, TestingCatalog reported. Early users named at launch include Yelp Host, Speak, Intercom’s Fin and Cognition’s Devin.
What to watch
The two launches are designed to work together: a voice front end that keeps talking while an agent does the work behind it. The concrete test is whether OpenAI adds data residency outside the US and ZDR support before the Agents API leaves beta. Those two gaps, not the price, decide whether regulated companies can move agent workloads onto it.
