AALLM
Get an API key
Cost & surface

Voice sessions and agent sandboxes: the non-token meter.

OpenAI’s September 10, 2026 releases added a voice endpoint billed per minute and a managed agent harness billed per container. Neither is token-metered, and neither fits a gateway route most teams already proxy.

Explore the live registry
New surfaces
2

GPT-Live 1 and the Agents API, both Sep 10, 2026

Voice session rate
$0.05 / min

billed per second, not rounded up

Routes on gpt-live-1
1 of 18

v1/live/sessions only — Responses and Realtime unsupported

Two new surfaces, and the meter each one carries

Observed September 14, 2026. The “separate” column is the usage that does not appear on the headline rate.

SurfaceRouteHeadline meterBilled separately
GPT-Live 1 · gpt-live-1v1/live/sessions$0.05 per minute, billed per secondBackend model and tool usage
Agents API (public beta)OpenAI-managed Codex harnessModel rates + tool rates + container ratesSandbox time for OpenAI-hosted environments
Claude Managed AgentsClaude APINo new unit — permission policyServer-side evaluation of each tool call

Sources: developers.openai.com · GPT-Live 1 model (rate and route) · developers.openai.com · Agents API overview (agent billing stack) · platform.claude.com · Release notes (permission policy)

The operating contract behind each release

What changes an integration, a cost model or an approval flow — not what changes a benchmark.

GPT-Live 1 — OpenAI, Sep 10, 2026

  • A full-duplex voice model for real-time conversations: it can listen and speak at the same time, and delegate reasoning and tool use to a backend agent.
  • Voice sessions cost $0.05 per minute, billed per second, and backend model and tool usage is billed separately. Session duration is not rounded up to the next whole minute — so the meter is continuous, not per-minute-blocks.
  • The model is reachable only on v1/live/sessions. Chat Completions, Responses, Realtime, Realtime translation and transcription, Batch, Fine-tuning, Embeddings and the audio and image endpoints are all listed as not supported for this model.
  • You choose the backend model or agent independently of the voice model, whereas the Realtime route uses one model for speech, reasoning and tool selection.
  • Your application owns permissions, confirmations, private function execution and durable task state, and interrupting speech does not automatically cancel backend work.
developers.openai.com · API changelogdevelopers.openai.com · GPT-Live 1 modeldevelopers.openai.com · Getting started with GPT-Live

Agents API — OpenAI, Sep 10, 2026 (public beta)

  • Gives an application access to the Codex harness through an OpenAI-managed API; OpenAI manages sessions, orchestration, context compaction and recovery.
  • Your application provides the tools and chooses the execution environment, and agents can operate in a sandbox where they execute code, edit files, connect to MCP servers and produce artifacts.
  • Billing stacks three units: model usage at the selected model’s API rates, OpenAI tools at their standard rates, and OpenAI-hosted sandboxes at standard container rates.
developers.openai.com · Agents API overviewdevelopers.openai.com · API changelog

Claude Managed Agents — Anthropic, Sep 10, 2026

  • Permission policies now include auto: the server evaluates each agent or MCP tool call and runs it, denies it, or pauses for your approval.
  • The agent.tool_use and agent.mcp_tool_use events report how each call was evaluated in an evaluation field alongside evaluated_permission, so an approval decision becomes auditable per call.
platform.claude.com · Release notes

Also official this week

Two lifecycle changes published between September 7 and 14, 2026 that alter how much notice a retirement gives you.

AWS Bedrock: a new lifecycle policy for models launched from September 7, 2026

  • The Bedrock lifecycle page now describes the policy for models launched on or after September 7, 2026, with models launched earlier following the legacy page.
  • Each model card carries an “EOL no sooner than” date plus its notice period. There are two Legacy periods, 6 months and 45 days, and most models have the 6-month period.
  • Once the Legacy period begins, new customers cannot adopt the model and existing customers may lose access after 15 days of inactivity — so an idle staging deployment can lose access before the published EOL date.
docs.aws.amazon.com · Bedrock model lifecycle

Microsoft Foundry: the replication delta on the newest image models

Microsoft’s retirement schedule already lists gpt-image-2.5-flare and gpt-image-2.5-sunburst with a 2026-09-09 version date and a 2027-09-09 retirement — one day after OpenAI announced the image models. The same schedule still carries gpt-realtime-2 (preview, 2026-05-06) at a 2026-08-31 retirement and gpt-realtime-mini (2025-10-06, GA) at 2026-09-21, which is the reminder that voice endpoints rotate faster than chat endpoints.

learn.microsoft.com · Model retirement schedule

What changes for a routing decision

  • A gateway that routes Chat Completions, Responses and Realtime has no route for voice. GPT-Live traffic needs an explicit v1/live/sessions path, plus WebRTC, WebSocket, sideband or SIP transport — this is not a model swap on an existing route.
  • Per-token budget alerts will not see a voice session. The session is metered in seconds while the delegated work is metered in tokens on a second meter; a spend cap expressed only in tokens governs neither the duration nor the backend calls it triggers.
  • Container time is now a budget line. Agent runs billed per container hour mean an unbounded sandbox can cost more than the model tokens it consumes, so cap sandbox lifetime and count container rates in the agent cost model.
  • The approval boundary differs by provider. Anthropic evaluates tool calls server-side under auto policies, while OpenAI leaves permissions, confirmations and function execution with the application — who is accountable for a tool call is a per-provider property, not a platform-wide one.
  • Preview voice endpoints carry short published lifecycles. Keep business-critical voice traffic on GA endpoints and track the per-model Bedrock card, where the notice period can be 45 days rather than 6 months.

Sources

All official pages observed September 14, 2026 (Europe/Paris). Claims are taken from the provider pages below; the live registry renders the deployed records separately.

Related live views: the September 2026 frontier wave and the OpenAI vs Azure AI deployment comparison render the current catalog records.

Deployment intelligence, explained

Why a second meter changes the routing decision

Deployment intelligence has assumed token pricing: an input rate and an output rate per million tokens, with cache reads as a discount. A session billed per second and a sandbox billed per container time break that assumption — the cost unit is now time or infrastructure, not only tokens.

The same week, agent harnesses gained their own governance surfaces. Read the endpoint table and the billing section together before routing voice or agent traffic: a model name does not tell you which route, which meter, or which approval boundary applies.