OpenAI turns the harness behind Codex into an Agents API

OpenAI has spent two years running long agent sessions in Codex and ChatGPT Work, and the parts that make those sessions survive for hours are now sold separately. The Agents API creates a production agent in one call that names the task, the model, the tools and the environment. It is in public beta for every developer, and it carries no fee of its own: the bill is the tokens and tools the agent consumes.

What the API actually hands over

Three pieces of plumbing that every agent team writes itself now arrive from the vendor. Context compaction runs automatically as a session approaches its context limit, so a workflow can span several context windows without custom code. Tool search loads tool definitions only when they are needed, which holds down token use and preserves the model's cache. Programmatic tool calling lets the agent run calls in parallel and filter results in code before anything returns to the context.

The harness underneath is the open-source Codex harness, so the logic that coordinates model calls, tools and context can be read in its public repository. OpenAI operates and maintains it, and ships versioned access to new harness capabilities alongside each model launch.

Where the agent runs is a choice

Compute is not bundled. An agent can run in a sandbox OpenAI provisions and manages, on the developer's own infrastructure, or with a sandbox partner; OpenAI names Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop and Vercel. That covers deployments inside a company VPC, specific file and secret storage, and CPU, GPU and memory profiles chosen for cold-start and cost.

• A session is created with client.beta.agents.sessions.create, which names the agent, its tools and its environment.
• Tools can be MCP servers, custom functions, or built-ins such as web search.
• Subagents run in parallel under multi_agent, each keeping its own context while the main agent combines the results.
• The OpenAI-hosted sandbox is the same infrastructure that runs Codex and ChatGPT.

Voice now has a per-minute price

GPT-Live-1, first seen in ChatGPT, reached the API on the same day at $0.05 per minute for the front-end voice layer. It listens and speaks in one model instead of chaining speech recognition, a language model and speech synthesis, and it hands reasoning and tool calls to a backend model that is billed separately. Developers shape tone, pace and conversational style through the system prompt, and the model supports telephony, so a full-duplex agent can answer a phone call.

The numbers published with it are the vendor's own. OpenAI reports a 30 percentage point gain on Full Duplex Bench over GPT-Realtime-2.1, and says the language app Speak measured interruptions falling by almost 80% against its previous turn-based system. No independent measurement of either figure exists yet.

The third release of the day

Alongside the two developer releases, OpenAI introduced a Data agent in ChatGPT Work. It connects to company data sources, investigates what changed, and builds shareable interactive dashboards from a question asked in plain language. Administrators decide which connections exist and which roles may use them, and queries run under the permissions of the connected account, down to table, row and column restrictions.