OpenAI Ships the Agents API in Public Beta and Brings GPT-Live-1 to the API

On September 10, 2026 OpenAI published two developer announcements: Introducing the Agents API and Build more natural voice experiences with GPT-Live-1 in the API. The first puts the agent harness and infrastructure behind Codex into public beta for all developers; the second moves ChatGPT's full-duplex voice model into the API at $0.05 per minute for the front-end voice layer.

Agents API: a production-ready agent in one call

OpenAI says a single API call that specifies the task, model, tools and environment creates a runnable agent; the model field in its sample code is gpt-6-astra. The highlights it lists:

• Environment choice: an OpenAI-managed sandbox, your own infrastructure, or sandbox partners — Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop and Vercel.
• Long sessions: earlier context is automatically compacted as a session approaches its context limit, so developers do not implement their own compaction logic.
• Tool use: tool search loads relevant tool definitions on demand to cut tokens and cost while preserving the model's cache, and programmatic tool calling lets agents run calls in parallel, chain operations and filter results in code. MCP, custom functions and built-in tools such as web search are supported.
• Subagents: with multi-agent support enabled, complex tasks are split and delegated to parallel subagents that keep their own context, with a configurable concurrency cap.
• Open-source foundation: the API runs on the open-source Codex harness, which OpenAI operates while developers can inspect the public codebase.

On billing, the post is explicit: "There are no additional fees for using the Agents API - you simply pay for the tokens and tools your agents use."

GPT-Live-1: one model listening and speaking at once

First introduced in ChatGPT, GPT-Live-1 listens and speaks at the same time and delegates deeper reasoning and tool calls to a backend text model such as GPT-6 Astra or a third-party model, replacing stitched-together speech-to-text, reasoning and text-to-speech pipelines. Capabilities OpenAI lists include interruption handling, shaping tone, pace and conversational style through the system prompt, silent context management under background noise and silence, long-session reliability, and telephony support. It natively provides ASR transcripts and response text, supports keyword biasing, and natively supports turn detection.

• OpenAI says that across its evaluations GPT-Live-1 improves Full Duplex Bench performance by 30 percentage points over GPT-Realtime-2.1, with large gains in turn-taking latency and interactive behavior, and that paired with GPT-6 Astra at medium reasoning effort it ranks first on Tau3, its measure of frontier voice-agent intelligence on end-to-end tasks.
• The post also reports that in early evaluations Speak found GPT-Live-1 gave learners more time to think, cutting interruptions by almost 80% versus previous turn-based systems.
• Voice options expand from a small set to a broader selection across accents, dialects and languages, with Quartz, Ripple, Vesper, Willow, Stone, Gleam, Meridian, Bossa, Tempo, Beacon, Delta and Cinder named, and more promised over the coming months.

The common thread

Both releases package infrastructure OpenAI already runs internally — the Codex harness and ChatGPT's full-duplex voice stack — as API products, with a clear cost story: no surcharge on the agent harness, and per-minute pricing for the voice front end. Customer quotes in the posts emphasize engineering simplification and turn-taking quality, with Yelp CTO Alex Levy saying GPT-Live-1 improved turn-taking and accuracy over their traditional voice architecture and lifted call handling rates. Pricing, concurrency limits and supported endpoints are specified in OpenAI's docs and pricing page.