Agents API, GPT-Live 1, and Decisions API
On this page
Models are only part of the fall announcements. For developers, three APIs change how you build: the Agents API, which hands orchestration off to OpenAI; GPT-Live 1, now available in the API for real-time voice; and the Decisions API, which turns text and images into typed responses. This lesson covers each based on the changelog and official announcements, then walks through the rest of the fall updates.
The Agents API: A Hosted Agent
Released in public beta on September 10, 2026, the Agents API lets you build agents on the same harness that runs Codex, maintained by OpenAI. OpenAI handles session orchestration, context compaction as a session nears its limit, and recovery after incidents. You work with durable sessions, which let you continue work over multiple turns and track progress in real time.
You create a session in a single call that describes the task, model, tools, and runtime environment. Tools can be MCP servers, your own functions, or built-in tools like web search; tool search loads definitions only when needed, and Programmatic Tool Calling is enabled by default (Lesson 5). A multi-agent option lets the primary agent delegate independent chunks to sub-agents that work in parallel, each with its own context. The agent runs in a sandbox hosted by OpenAI, on your own infrastructure, or with a partner: OpenAI specifically cites Cloudflare, E2B, Modal, and Vercel.
Several additions followed within a few weeks. On September 29, the Agents API received computer use: the agent performs tasks in a browser hosted by OpenAI, while your application manages site approval and login. On October 9, you could pre-warm a sandbox before attaching a session to it, reducing user wait time, and billing for hosted sandboxes now applies only during active turns. On October 16, these sandboxes integrated skills to create and edit spreadsheets, presentations, and Word documents, and the Agents API received Ultrafast mode.
For anyone building with Agent Builder, this is important: OpenAI is shutting down Agent Builder on November 30, 2026, and directing users to the Agents SDK or ChatGPT workspace agents. Lesson 12 details this timeline.
GPT-Live 1: Full-Duplex Voice in the API
In July, GPT-Live rebuilt ChatGPT’s voice, but the API wasn’t open yet. That changed on September 10, 2026: GPT-Live 1 is generally available. It lets you build full-duplex voice conversations that continue while a model or agent in the background reasons and calls tools. Two delegation modes exist: to an OpenAI model via the Responses API, such as GPT-6 Astra, or to your own backend, which can rely on another provider. The voice session costs $0.05 per minute, billed by the second; the background model and tools are billed separately.
On October 16, OpenAI opened voice creation from a text description to all API customers, free to create, and added nine natural voices, including two in French: Brise and Sillage. For transcription, two models launched on July 28: gpt-transcribe, for files, and gpt-live-transcribe, for real-time streaming. They replace whisper-1 and the gpt-4o-transcribe family, which are scheduled for deprecation on February 26, 2027.
The Decisions API: Typed Responses, Fast
The Decisions API, launched in beta on October 6 and generally available since October 22 with GPT-6 Luna, addresses a very common need: asking a model a series of closed-ended questions about text, an image, or structured data, and getting typed responses. OpenAI reports decisions up to ten times faster than calling GPT-6 Luna via the Responses API. Since general availability, a request can contain up to 400 questions, up from 200 during beta; common inputs can be cached, images can appear in a question’s instructions, and structured data passes directly via input_json. Since October 15, it also integrates with GPT-Live: a voice application can act while the user speaks.
Think of it for ticket triage, routing requests, verifying document compliance, or extracting a few fields from thousands of documents: where you used to write a classification prompt and an output schema, a set of typed questions suffices.
Other Fall Updates
| Date | Update |
|---|---|
| September 8, 2026 | GPT Image 2.5 Sunburst and Flare, for image generation and editing |
| October 6, 2026 | API usage tiers reduced from five to three: Build, Launch, and Grow |
| October 20, 2026 | Prompt caching extended to 12 hours |
| October 21, 2026 | Project-by-project selection of allowed service levels |
| October 28, 2026 | Up to $500 in free API credits per month for eligible ChatGPT Pro and Business subscribers |
| October 30, 2026 | GPT-6 Rosalind, for life sciences research, via approved access program |
| November 4, 2026 | GPT-Image-2.6 Sunburst and Flare: adjustable complexity, SVG and animated GIF generation |
| November 5, 2026 | GPT-6-Cyber, for vulnerability research in authorized environments, upon approval |
Your Turn
Choose the right API. Pick three needs in your organization: for example, an agent that prepares a daily incident summary, a phone system that routes calls, and triage of 5,000 incoming emails per day. For each, state which API you would try first—the Agents API, GPT-Live 1, the Decisions API, or the Responses API alone—and why. Note what you would measure to confirm that choice.
Official sources: OpenAI API Changelog · GPT-Live 1 (OpenAI API documentation) · Introducing the Agents API · GPT-Live 1 in the API · Decisions API (OpenAI)
Key Takeaways
- The Agents API (public beta, September 10) hosts agent orchestration: durable sessions, compaction, sub-agents, sandboxes, computer use
- GPT-Live 1 has been available in the API since September 10 at $0.05 per minute of voice session; two French voices added October 16
gpt-transcribeandgpt-live-transcribereplace older transcription models before February 26, 2027- The Decisions API (available since October 22) produces typed responses at scale with GPT-6 Luna