Developers can keep execution on their own infrastructure while OpenAI operates orchestration and bills for model and tool use. The design reduces infrastructure lock-in; the remaining diligence issue is whether managed operating state can move as cleanly as the sandbox.
OpenAI opened the Agents API to public beta on September 10 with a split architecture that matters more to a CIO than another model benchmark. OpenAI runs the agent harness, while developers can execute code in an OpenAI sandbox, on their own infrastructure or through a partner. There is no separate Agents API platform fee in beta; revenue comes from the models and tools the agent consumes.
That bargain lowers one large deployment cost without forcing the customer's execution environment into OpenAI's cloud. Files, repositories, code execution and network access can remain under enterprise control while OpenAI maintains context management, tool routing and other model-facing orchestration. The commercial trade is lower harness-engineering burden in exchange for recurring OpenAI consumption and reliance on managed behavior above the sandbox.
OpenAI had already made execution portability an explicit design goal. Its April Agents SDK let developers bring their own sandbox, describe a workspace through a portable Manifest abstraction and externalize state so a failed or expired sandbox could be restored into a fresh environment. September moves more of the orchestration from a toolkit the customer assembles into a service OpenAI operates.
Execution portability is documented; managed-state portability is less explicit
The April SDK documentation gives developers concrete mechanisms for moving the working environment. Workspaces can mount local files, define output directories and connect external object stores including AWS S3, Google Cloud Storage, Azure Blob Storage and Cloudflare R2. The same workspace abstraction can be used across supported sandbox providers.
The Agents API adds capabilities above that layer: automatic context compaction for long sessions, tool search, parallel tool calls and subagent coordination. Those services can affect how an application behaves even when its compute remains customer-owned.
The September launch material is explicit about execution portability and silent on a general export mechanism covering every managed long-session object, harness setting, evaluation artifact or governance configuration. Silence on those objects is an evidence boundary rather than proof of non-portability. OpenAI has demonstrated how the execution environment can move; the public record is less complete on how much managed operating state can be reproduced outside the service.
For platform selection, that changes the diligence question. The risk is not that OpenAI must own the servers. It is that a production workflow may come to depend on service-level behavior whose migration path is less documented than the sandbox itself.
Financial Services adds data and governance to the same decision
ChatGPT for Financial Services extends that architecture into a higher-value workflow. OpenAI combines GPT-6 Astra with finance-specific tools and data relationships. Morgan Stanley and Evercore were design partners. Daloopa, PitchBook, LSEG News and Crunchbase are presented as built-in data sources, while firms can connect existing entitlements from providers including S&P Capital IQ, MSCI, Dow Jones Factiva, Moody's and LSEG.
The two distribution models have different switching economics. Entitlement integrations preserve the institution's direct commercial relationship with the underlying data vendor. Data indexed and hosted inside the OpenAI product is more closely tied to the retrieval and workflow layer in which employees use it.
Administrators can also configure firm templates, app and skill permissions, retention rules, compliance logging and separated workspaces. Those features make the product more useful precisely because they encode operating policy around the model. The launch establishes the controls; their reproduction cost elsewhere is undisclosed.
OpenAI's enterprise position is stronger because customers retain infrastructure control. Customer-run compute removes a major adoption objection while the company concentrates its product differentiation in orchestration, authorized data access and governance. For CIOs, the unresolved architectural term is now specific: how much production state and policy can be exported or recreated independently if the managed harness is replaced. The current documentation leaves that question open, making it an architecture diligence issue rather than evidence of inevitable lock-in.
Sources
OpenAI, Introducing the Agents API: https://openai.com/index/introducing-the-agents-api/
OpenAI, The next evolution of the Agents SDK: https://openai.com/index/the-next-evolution-of-the-agents-sdk/
OpenAI, Introducing ChatGPT for Financial Services: https://openai.com/index/introducing-chatgpt-financial-services/
Reuters via Investing.com, Agents API launch coverage: https://ng.investing.com/news/stock-market-news/openai-launches-agents-api-in-public-beta-for-developers-93CH-2692209