OpenAI Rents Out the Codex Harness as a Managed API

OpenAI opened its Agents API to public beta on Sept. 10, packaging the same Codex harness that runs its own coding agent into a managed service that handles session memory, context compaction and crash recovery for agents that run for hours at a stretch.

The API organizes work around four objects, MarkTechPost reported: an agent, defined by its model, instructions, tools and MCP servers; an optional sandboxed environment for file access and command execution; a durable session that persists across a long task; and the events and items that make up its inputs and outputs. OpenAI said the harness automatically compacts earlier context as a session nears its limit, reduces token use through tool search, and supports subagents that delegate parts of a task to separate, independently scoped contexts.

Nine sandbox vendors signed on before launch

Developers can run agents in OpenAI-hosted or self-hosted sandboxes, or through nine first-class integrations named at launch: Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop and Vercel, according to MarkTechPost's review of the announcement. OpenAI is not charging a separate fee for the API itself; customers pay only for model tokens, tool calls and any container time they consume, and the service keeps data in the U.S. only during the beta, with no support yet for Zero Data Retention accounts.

OpenAI's own customer figures, which The Terminal could not independently verify, claim strong early results: MarkTechPost reported that OpenAI cited a jump from a 0.71 to a 0.85 evaluation score and a fourfold latency cut at a company called Ciridae, a 60% cost reduction per case at SafetyKit, and an 86% drop in failed responses at Hypha. Those numbers are vendor-supplied and describe customers OpenAI selected to publicize, not an independent audit.

A voice model shipped the same week, at a fixed per-minute price

GPT-Live-1, the full-duplex voice model OpenAI has run inside ChatGPT since July 8, became directly callable through the API the same week, at $0.05 a minute, Gekro reported. The launches point at the same strategy: OpenAI is converting capabilities that once lived only inside its own consumer app, coding agent and now voice interface into metered infrastructure other companies build on, mirroring how it turned the Habitat storage system behind ChatGPT into public numbers rather than an internal detail.

Product	Status before Sept. 10	Status after Sept. 10
Codex harness	Internal to OpenAI's own coding agent	Rentable via the Agents API, public beta
GPT-Live-1	ChatGPT-only since July 8	Callable via API at $0.05 a minute

The Agents API arrives as rivals chase the same long-running-agent category from the tooling side rather than the model side: Cognition, maker of the Devin coding agent, raised $2 billion at a $48 billion valuation on Sept. 8, a bet that the harness and orchestration layer around a model, not just the model, is where the value sits. OpenAI's move to sell that layer directly, rather than leaving it to third-party frameworks, puts it in competition with some of the same startups building on its own models.

OpenAI has not said when the Agents API will exit public beta or what pricing changes might come with general availability. Whether the nine sandbox partners see meaningful usage, rather than a launch-day mention, will show up in their own developer numbers before OpenAI publishes anything further.