News · 2026-09-10

OpenAI turns the Codex harness into a product with an Agents API

OpenAI released an Agents API in public beta on 10 September 2026, letting developers build on the same managed Codex harness that powers its own Codex product. According to OpenAI's API changelog, the service handles "session orchestration, context compaction, and recovery" on OpenAI's side, with durable sessions that persist across turns, streamed progress, custom tool and MCP server connections, and sandbox options. The same entry made GPT-Live 1 generally available at five cents a minute.

Key facts

• The Agents API entered public beta on 10 September 2026
• GPT-Live 1 reached general availability the same day at $0.05 per minute, billed per second
• The underlying Codex harness was open-sourced under Apache 2.0 in August 2026
• Primary source: OpenAI API changelog

The harness is the product now

For most of the past two years the model was the product and everything around it was plumbing you wrote yourself. That plumbing turns out to be where a great deal of the capability lives.

A harness is the loop around a language model: it calls a tool, reads what comes back, decides what to do next, keeps track of what has happened, and handles the many ways this goes wrong. Anyone who has built one knows that the model is the easy part. The hard parts are what happens when a session outgrows its context window, when a tool call hangs, or when work needs to survive a process restart.

OpenAI is now selling exactly those three things. Context compaction addresses the first, recovery the second, durable sessions the third.

The strategic reading is that OpenAI has decided the scaffolding is defensible. It open-sourced the Codex harness in August under an Apache 2.0 licence, which looked at the time like giving away the crown jewels. What it was actually doing is clearer now: give away the specification, sell the operated version. The same harness reportedly powers both Codex and ChatGPT Work internally, so what developers get is the thing OpenAI runs for itself.

Why this arrives now

It lands in a week that has made the same argument three times over from different directions.

Cognition's SWE-2 launch is a coding product built by post-training somebody else's open model — the differentiation is in the training and the agent loop, not the pretraining. A robotics paper published the day before argues that a general vision-language model can control robots competently given the right interface, without any new model capacity. And Ground Truth has covered the failure side repeatedly, from agents that build their own harness and never once save state to seven live agents sending $12,431 in unsolicited invoices.

The through-line is that agent behaviour is substantially determined by the software wrapped around the model. That cuts both ways: a well-built harness gets more out of a weaker model, and a badly built one lets a strong model do real damage. Buying the harness from the lab that built the model is a reasonable response, with the obvious cost of depending on that lab for the parts you can no longer inspect.

The voice half

The same changelog entry moved GPT-Live 1 to general availability, for "full-duplex voice conversations that can continue while a backend model or agent handles reasoning and tools."

Full-duplex means both parties can speak at once — the model keeps listening while it talks, so it can be interrupted like a person rather than waiting for silence. The architectural detail worth noting is the split: the voice model runs the conversation while a separate model does the thinking behind it. That is why the pricing is per minute of conversation rather than per token, and it is a sensible division of labour, since the responsiveness a voice needs and the deliberation a hard question needs are different problems.

Five cents a minute works out to three dollars an hour. Whether that is cheap depends entirely on what the conversation is worth, but it is comfortably below the cost of a human doing the same talking.

The honest caveat

This is a changelog entry, not a technical report. There is no published benchmark showing that the managed harness outperforms a well-built custom one, no stated pricing for the Agents API itself in the entry, and no service-level commitments in what OpenAI has published so far. Public beta means the interface can change.

There is also a dependency question that developers should weigh rather than assume. Handing session state, context management and recovery to a provider means that when the agent behaves strangely, the part you most need to debug is the part you cannot see. That trade is familiar from every managed service, but it is sharper here because agent failures tend to be emergent and hard to reproduce — and, as this week's disclosure from Anthropic showed, an agent's own account of what it was doing is not always reliable evidence.

Primary source, verified: read the paper →

Ground Truth. (2026, September 10). OpenAI turns the Codex harness into a product with an Agents API. Ground Truth. https://groundtruth.day/news/openai-turns-the-codex-harness-into-a-product-with-an-agents-api.html

@misc{groundtruth:openai-turns-the-codex-harness-into-a-product-with-an-agents-api,
  title  = {OpenAI turns the Codex harness into a product with an Agents API},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {sep},
  url    = {https://groundtruth.day/news/openai-turns-the-codex-harness-into-a-product-with-an-agents-api.html}
}

Topics: openai · agents · api · developer-tools · harness · product-launch