Public beta is the gate. OpenAI posted the Agents API on 10 September 2026: the Codex harness as a managed API, not a new model SKU.
The news
OpenAI's 10 September 2026 product post introduces the Agents API in public beta. The company says the API brings the same harness and infrastructure that powers Codex to developers: context management, tools, subagent coordination, and environments that keep long-running agents working with files, code, and intermediate results for days.
You create a session by specifying task, model, tools, and environment. The sample on the page uses model gpt-6-astra, MCP tools, multi-agent with max concurrent subagents set to 3, vault IDs, and an openai_hosted environment with capability directories. OpenAI hosts and maintains the harness. You choose compute: an OpenAI-managed sandbox, your own infrastructure, or a sandbox partner.
Named partner integrations on the post: Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel. Options listed: fully managed environments or VPC deployments; file and secret storage choices; different CPU, GPU, and memory profiles.
Harness features named: automatic context compaction as a session approaches its limit; tool search that loads tool definitions as needed; programmatic tool calling for parallel and chained tool use in code; MCP, custom functions, and built-in tools including web search; multi-agent support that delegates independent pieces to subagents with their own context.
The Agents API is powered by the open-source Codex harness. OpenAI points to the public openai/codex codebase for visibility into coordination logic while OpenAI operates the managed path.
Pricing line on the post: no additional fees for using the Agents API; you pay for tokens and tools on the pricing page. Available in public beta to all developers. GA date is UNKNOWN.
Customer quotes on the page are company-attributed product claims. Ciridae's CTO says an evaluation score moved from 0.71 to 0.85 and claims 4x latency reduction on subagent flows. SafetyKit claims a 60% reduction in cost per case after migration. Hypha claims an 86% reduction in failed agent responses by separating harness from sandbox. Those figures are not The Frontier's measurements.
Who is bound
OpenAI ships the API and the managed harness. Developers on the public beta are the audience. Sandbox partners listed above integrate environments. End customers of Ciridae, Long Lake, WithCoverage, SafetyKit, Dwelly, Hypha, deepsense.ai, and Nash.ai appear only as named design partners or quote sources on this page, not as contractual counterparties of the reader.
What's new
A public-beta Agents API path that externalizes the Codex-class harness: sessions, hosted or partner sandboxes, multi-agent, tool search, and compaction without building that stack yourself. It is not GPT-6 Astra's 3 September model launch. It is not the separate GPT-Live-1 voice SKU posted the same day. It is not the older Agents SDK evolution posts from April 2026.
What it does not settle
Date: 10 September 2026. Status: public beta; general availability date UNKNOWN. Price: no separate Agents API fee stated; token and tool rates sit on the pricing page and are not restated here. Region and data residency for hosted sandboxes are UNKNOWN on this page. SLA, retention, and audit log detail are UNKNOWN here. Whether every Codex CLI feature maps one-to-one is UNKNOWN. Customer percentage claims stay tagged to those companies.
What to do now
Platform teams already running self-hosted agent loops should read the Agents API overview and the environment docs before ripping orchestration. Security should treat partner sandboxes and OpenAI-hosted sandboxes as different trust boundaries. Ignore the customer score quotes as your own baseline until you reproduce them. If you only needed a chat completion, this post is the wrong tab.