OpenAI Opens Agents API Beta, Puts Codex Harness Behind One Call

OpenAI opened a public beta of its Agents API on Thursday, giving developers a single endpoint to run its Codex coding-agent harness instead of building their own orchestration layer around the model. The API runs on gpt-6-astra and requires an OpenAI-Beta: agents=v1 header on every request, according to OpenAI's own documentation.

OpenAI said the beta carries no fee beyond what a developer already pays for the model: standard token rates, standard tool rates, and standard container rates when an agent runs inside an OpenAI-hosted sandbox.

What the four new primitives do

The documentation defines four objects. An Agent holds the model, its instructions, its available tools and any MCP server connections, a protocol that lets a model call external data and tool servers. An Environment is the sandbox or infrastructure the agent can reach. A Session is a persistent agent instance that keeps state across many turns. Events and items record what the user sent and what the agent produced.

OpenAI said it now runs "sessions, orchestration, context compaction, and recovery" on its own infrastructure, work a developer previously had to write.

Where it sits next to two other OpenAI APIs

OpenAI now offers three ways to build an agent. The Agents SDK runs inside a developer's own application. The Responses API is an application with optional hosted orchestration. The new Agents API instead runs the full harness on OpenAI's own infrastructure, with the option to point it at nine partner sandboxes instead of an OpenAI-hosted one: Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop and Vercel, MarkTechPost reported.

What the beta does not support

The Agents API "currently supports data residency only in the United States and does not support Zero Data Retention (ZDR)," OpenAI's documentation states. Zero Data Retention is the setting some regulated customers require so that OpenAI does not store request and response content after answering it. OpenAI did not say when, or whether, either limitation will lift.

Numbers OpenAI is citing from early customers

MarkTechPost's writeup of OpenAI's release cites three vendor-supplied results, not independently verified:

1. SafetyKit cost cut60 percent change
2. Hypha failed-response cut86 percent change

Source: OpenAI, via MarkTechPost, Sept. 10, 2026

A fourth figure does not fit the chart's percentage format: Ciridae told OpenAI its evaluation score rose from 0.71 to 0.85 with a fourfold cut in latency. None of the three companies' methodology is described beyond OpenAI's own account of it.

What came three weeks earlier

OpenAI's Assistants API reached its sunset on Aug. 26, 2026, a year to the day after OpenAI announced that date, and the company pointed developers to the Responses and Conversations APIs instead, according to OpenAI's deprecations page. The Agents API beta arrived 15 days later, adding a fourth interface to a lineup OpenAI has now told developers to consolidate around twice within six months. Two prior OpenAI releases show the pattern this fits: the company priced its GPT-Live-1 voice model by the minute in the same API family, and it built the new agents around gpt-6-astra, the model OpenAI ships behind its own capability gate.

OpenAI did not say when the beta will reach general availability, or when data residency outside the United States might arrive.