OpenAI Launches Agents API in Public Beta with Codex Tools

OpenAI has launched the Agents API in public beta on September 10. This is the same harness that powers Codex and ChatGPT in agent mode. From now on, any developer with a paid account can orchestrate sessions lasting days, with automatically compacted context, attached MCP tools, and three types of execution sandbox: OpenAI infrastructure, own server, or partner. Cloudflare, DigitalOcean, and Oracle joined as environment partners on the launch day. The pricing excludes additional fees; the company only charges for the token consumption of the chosen model.

The API structure has been streamlined to four objects. The Agent defines the model, instructions, and tools. The Environment describes the optional sandbox. The Session maintains persistent state and the event stream. Each Session can generate subagents that inherit part of the parent context, with automatic compaction when the window saturates. This is the same internal design that OpenAI uses to run the Codex CLI, now made available to the public on equal footing.

Direct Challenge to Anthropic and Google

This move puts pressure directly on Anthropic, which opened the Claude Agent SDK as an open framework in July without offering a managed sandbox, and Google, whose Vertex AI Agent Builder charges an orchestration layer in addition to token consumption. The obvious takeaway is pricing: for the first time, a frontier company delivers long-duration sessions without markup on the infrastructure. ChatGPT in agent mode had been operating with this harness since March, but only for consumer products.

The less obvious takeaway is distribution. By partnering with Cloudflare, OpenAI gains reach in lower latency regions that Azure does not adequately cover, especially in Latin America and Southeast Asia. With DigitalOcean, it targets solo developers and startups avoiding exit costs from AWS. With Oracle, it positions the Agents API within the $300 billion Stargate commitment announced in July, ensuring that consumption boosts OCI's own RPO, which closed Q1 of fiscal year 2027 at $664 billion, according to a report released on Thursday.

Where the Beta Lands Outside the U.S.

In the UK and Germany, the Agents API enters a denser regulatory environment: the European AI Act, in effect since August 2026 for general use systems, requires sandbox providers to document agent decision tracking. OpenAI has yet to publish the log model for compliance, and European integrators, including Capgemini and Atos, have already indicated that implementation for financial clients will require an additional audit layer. Deutsche Bank, which runs 1,400 production use cases with a proprietary framework, said last week that it will conduct a restricted pilot before any migration.

In India, where OpenAI's developer base grew to 3.2 million in August according to company data, the Agents API directly collides with vertical agent providers created by TCS (Ignio Agentic), Infosys (Topaz), and Wipro (HOLMES Agents). The pain point is not technical; it’s commercial. A practice that previously sold $200,000 per customized agent project now competes with an SDK that any junior engineer can assemble in three days. Analysts at HFS Research project that this contract range will see a repricing of about 40% by December.

There are skeptics. For Aiman Ezzat, CEO of Capgemini, the commoditization of the harness does not eliminate the true barrier to entry, which is vertical knowledge and integration with legacy systems. The counterpoint is backed by data: even with Microsoft’s Copilot running since 2023 in SAP, Salesforce, and Oracle, integrators still charge an average of $850,000 for corporate deployment in Fortune 500, according to the latest survey by Constellation Research published in July.

Codex originally aimed to be a coding assistant; the new API applies this logic to any business process lasting more than a few hours. This echoes what Anthropic has been arguing with Claude Code, now with the difference that OpenAI covers the machine costs. If the model gains traction, the next competition will be about who controls the operating system for agents, rather than who has the best isolated model.