OpenAI has transitioned its Agents API to public beta on September 10, 2026, thereby opening the infrastructure behind the coding agent Codex for any applications. Developers gain session management, memory cleanup, and tool control through a single programming call, without having to build these components themselves. Early customers report significantly lower costs and reduced error rates.
Interface consolidates session management and tool control
The Agents API is based on the Codex-Harness, that control layer which allows a language model to plan multi-step processes, call tools, and evaluate intermediate results, rather than just answering single requests. OpenAI takes on tasks that developers previously had to program themselves: sessions are maintained across multiple requests, long conversation histories are automatically compressed, and agents can delegate sub-tasks to other agents. Additionally, the interface automatically searches for available tools as needed, so developers no longer have to list every possible function in the prompt beforehand. For executing code or file operations, OpenAI offers a hosted sandbox that uses the same infrastructure as Codex and ChatGPT.
The foundation for this is the same Codex technology that OpenAI previously expanded with a Persistent Mode for permanently running agents and which has been encrypting messages between agents since June 2026. OpenAI does not charge additional costs for the interface itself: billing is solely based on used tokens as well as tool and container consumption, as with the previous API.
Early customers report lower costs
Three companies mentioned by OpenAI are said to have already benefited from the transition. The security service provider SafetyKit, which assists companies in complying with content guidelines, has reduced its operating costs per processed task by 60 percent because the Agents API handles retries and caching itself. The automation provider Hypha, which manages support processes for corporate clients, reports 86 percent fewer failed agent runs after switching from a self-built orchestration to the managed interface. The startup Cirridae, specialized in agent backends, reportedly achieves a fourfold lower response time according to OpenAI.
These figures are not independently verified – they come solely from OpenAI’s own announcement. For companies operating their own support or automation bots, such effects potentially reduce the effort required to operate AI agents in production reliably, without the need to build a dedicated team for operations and troubleshooting. Until now, developer teams usually had to build orchestration, error handling, and sandbox infrastructure from scratch, which typically took weeks to months of development time and then incurred ongoing maintenance efforts.
Nine partner sandboxes complement the own offering
Those who do not want to run their agent code in OpenAI’s own sandbox can choose between their own infrastructure and nine partner environments: Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel. All connect via the same Agents API, allowing developers to switch the execution environment with a single parameter change without adjusting the rest of the code. For companies with their own data storage requirements or existing contracts with one of the partners, this eliminates the need to rely solely on OpenAI’s own infrastructure. Which option is worthwhile depends primarily on how much control over network and storage locations a team needs.
Access to the Agents API is granted to developers through their existing OpenAI Platform account; a separate access route or waiting list is not provided. OpenAI does not specify any regional restrictions for the new interface; the EU data residency that OpenAI had previously introduced for its API offering continues to apply unchanged, according to its own statements. Whether companies in Germany will receive the same conditions as in the USA remains open in the announcement.
It will be crucial whether the Agents API actually saves development time or whether companies will remain more tied to OpenAI’s infrastructure in the future, as sandbox and orchestration are hardly interchangeable outside of the nine partners. Competing with Google, Anthropic, and a growing number of open agent frameworks, OpenAI is targeting the same developer teams that have previously built their own solutions. A next milestone will be when OpenAI exits the beta phase and whether additional limits or pricing tiers will be added – the company has not yet provided a date for this.
