Ope​nAI's September Platform Day: GPT-Live-1 Reaches the API as the Agents API Opens in Beta

Two things left Ope​nAI's platform on September 10, 2026, and the quieter one has the larger blast radius. GPT-Live-1 — the full-duplex voice model that has been running inside ChatGPT since July 8 — is now callable by developers at $0.05 per minute of voice. Alongside it, and mentioned far less in the coverage, the Agents API entered public beta: the same harness that runs Codex, exposed as a hosted product with managed sandboxes. One is a better voice. The other is Ope​nAI selling the thing that used to be the hard part of building an agent.

Both were read off the primary sources for this piece: Ope​nAI's Agents API announcement, its developer-community post introducing GPT-Live-1 in the API, and the developer documentation for both. Where a number comes from Ope​nAI rather than an outside evaluator, it is labelled as such below — and one figure does come from outside, which changes the story slightly.

Artificial Analysis began publishing a Speech to Speech Index this year, and GPT-Live-1 has a row on it: 69.8%, a weighted average across speech reasoning, agentic performance, arena preference and task success. That places it seventh of the eleven models scored — behind GPT-Realtime-2.1 at 73.9%, which is the model it replaces, and behind Gr​ok Voice Think Fast 2.0 at 79.0%. The independent scoreboard and Ope​nAI's own benchmark table do not tell the same story, and both are true.

What shipped, in the order it will change your architecture

• Voice — GPT-Live-1 is live in the API as gpt-live-1, billed at $0.05 per minute of voice, charged by the second, with the backend reasoning model billed separately at its own rates.

• Agents — the Agents API is in public beta. Ope​nAI says there is no additional fee for the API itself; you pay for model tokens, tools, and hosted sandbox container time.

• Sandboxes — Ope​nAI now hosts the execution environment, reusing the same sandboxing infrastructure as Codex and ChatGPT, configurable with files, packages, skills and plugins.

• Environments — besides Ope​nAI's own sandbox, teams can point agents at their own infrastructure or at third-party sandbox vendors in the beta partner list.

The voice release is a model story. The Agents API is a platform story, and platform stories are the ones that quietly re-point a codebase.

The Agents API: an agent in one POST

The core call is POST /v1/agents/sessions, sent with an Ope​nAI-Beta: agents=v1 header, and the pitch is that task, model, tools and environment are enough to get a running agent back. SDKs cover Python, TypeScript/JavaScript, Go, Java and Ruby.

What makes it more than a wrapper is that Ope​nAI is operating the harness rather than shipping it. The AgentKit harness is open source and inspectable, but the running copy is Ope​nAI's — context compaction across long sessions, tool search, programmatic tool calling, MCP support, custom functions, built-in web search, and subagent orchestration with configurable concurrency. Versioned harness capabilities land alongside model launches, which is the interesting commitment: the scaffolding moves at the same cadence as the models.

For anyone who has spent a quarter maintaining a loop — retry logic, context trimming, tool routing, the subagent fan-out that always leaks state — that is the actual product. Not the model call. The plumbing around it that you no longer write.

Hosted sandboxes are the part with a bill attached

Agents that execute code need somewhere to execute it, and the beta ships Ope​nAI's own answer: a managed sandbox that runs code, works with files and produces artifacts, configurable with packages, skills and plugins. Developers who already have a hardened execution environment are not forced onto it — bring-your-own-infrastructure and a set of named sandbox partners are both in the beta.

Two operational caveats are worth filing before you build on it. Data residency in the beta is US-only, and Zero Data Retention is not supported. For a regulated workload, those two lines decide whether this is a pilot or a production path, and they are the details that tend to be discovered late.

GPT-Live-1 in the API: flat per-minute billing is the real change

The voice model itself is not new — it replaced Advanced Voice Mode inside ChatGPT on July 8 — but the commercial shape is. Under the Realtime line, audio was metered in tokens, which meant forecasting a call required modelling audio burn: how many tokens a second of silence costs, how a caller talking over the agent changes output length. A flat $0.05 per voice minute collapses that to multiplication. An hour of continuous open line is $3.00. A four-minute average across a thousand calls a day is roughly $200 a day.

Sessions run from a Live Sessions endpoint with a single model ID and no dated snapshots to pin. The documentation covers WebRTC, WebSockets and telephony/SIP as connection paths, and the published quickstart builds the WebRTC one. Billing has two meters, and only one of them is the voice: every delegated reasoning call, tool invocation and web search bills at the backend model's normal rates.

The architecture is delegation. GPT-Live-1 handles the conversation — deciding many times a second whether to keep listening, pause, interrupt or hand work off — and passes anything that needs real reasoning across an asynchronous boundary to a separate backend, which keeps working while the spoken conversation continues. The docs define two modes: Responses delegation, where Ope​nAI calls a supported Responses model for you and supplies conversation context, and client delegation, where your own application supplies the backend and keeps control of execution, context and which results reach the voice layer. In the published Responses example, that backend is GPT-5.6 Terra, named in a delegation object alongside its own instructions and tools. At the July consumer launch it was GPT-5.5; community write-ups since have pointed at GPT-6 Astra.

Every headline number for the voice layer is Ope​nAI's own and, at the time of writing, not independently reproduced by anyone: 80.1% on Full Duplex Bench v1.5 interactivity against 45.4% for GPT-Realtime-2.1, 0.798-second turn-taking latency against 1.41 seconds, 87% tool-calling success, and a 32% pass rate on a banking voice-support evaluation against 12.4% for the model it replaces. That last pair is the honest one — a large improvement, and still a system failing roughly two thirds of a regulated workflow. Third-party customer evidence exists but is thin and self-interested: Yelp for phone reservations, EliseAI, and learning platform Speak reporting close to 80% fewer interruptions than its previous turn-based stack.

The independent result is less flattering, and worth putting next to the list above rather than under it. On the Artificial Analysis Speech to Speech Index — one composite score over speech reasoning, agentic performance, arena preference and task success — GPT-Live-1 sits at 69.8%, below GPT-Realtime-2.1's 73.9% and well below Gr​ok Voice Think Fast 2.0's 79.0%. Read the two together and the shape of the product becomes clear: Ope​nAI built the best conversational mechanics available and did not build the best voice agent, because the reasoning is delegated to a model the index scores separately.

The two announcements are one announcement

Read together, the releases describe the same reorganisation from two directions. The Agents API moves the loop, the sandbox and the context management into a hosted product. GPT-Live-1 moves the conversational surface into a thin front end that delegates its thinking elsewhere. In both, the model stops being the thing you build around and becomes a component you select — and in both, the selection is a configuration line rather than an architectural commitment.

That is exactly the seam a routing layer sits in. OrcaRouter does not carry gpt-live-1: the Live Sessions endpoint is Ope​nAI's alone, and we route none of it. The delegation half is a different story. The backend the voice layer hands work to speaks the Responses API, and that is a normal model call — the model Ope​nAI's own example names, GPT-5.6 Terra, is available through OrcaRouter at Ope​nAI's list price of $2.00 per million input tokens and $12.00 per million output, with the Responses endpoint exposed. Client delegation goes further: it lets your application supply the backend outright, which means the model behind the voice can be any endpoint you control. The same is true of the Agents API's model slot: the harness is Ope​nAI's, but the model inside it is a choice you keep, and roughly 190 models from eleven upstream providers sit behind one key at provider list price with no markup on top.

Concretely, three things follow from that. Backends can be A/B'd on live traffic by editing one line, which is how you find out whether a cheap reasoner is good enough before committing a phone line to it. Failover can sit under the delegation model, so a bad afternoon upstream does not take the conversation down with it. And a task can be run against several backends in one call through the routing DSL, or answered by a panel of models at once with model fusion — useful the moment an agent's output has to be trusted rather than merely generated.

What is in neither release

• No image or video input on GPT-Live-1 — the multimodal voice surface older ChatGPT modes hold is still unaddressed.

• No structured outputs on the voice layer, so schema-validated tool arguments have to be enforced in the backend model.

• No Free-tier access to GPT-Live-1, and rate limits are denominated in concurrent sessions — 25 on Tier 1, rising to 500 by Tier 5.

• No Zero Data Retention and no data residency outside the US in the Agents API beta.

• No independent reproduction of any GPT-Live-1 benchmark figure.

The bet Ope​nAI made on September 10 is that developers want to buy the harness and choose the brain separately. The first half of that is now a line item. The second half is where the next twelve months of competitive pressure will land — because a hosted harness with a swappable model slot is only as good as the models you can put in it, and right now that is the one part of the stack Ope​nAI does not control alone.

The delegation half is a different story — the backend the voice layer hands work to is a normal model call, and GPT-5.6 Terra is available through OrcaRouter at Ope​nAI's list price with the Responses endpoint exposed.