AALLM
Get an API key
Cost & surface

Cost control for AI agents,
beyond the token meter.

Token budgets governed model calls. Voice sessions are billed per second, and agent sandboxes per container minute — neither appears on a token meter. ALLM normalizes the deployment-level meters that decide what an agent actually costs.

Get an API key Read the source report

The shift

The bill stopped being measured in tokens alone.

Three surfaces released in September 2026 carry a meter that a token rate does not describe. All figures observed from provider documentation on September 14, 2026.

SurfaceRouteHeadline meterBilled separately
GPT-Live 1 — real-time voicev1/live/sessions$0.05 / minute, billed per secondBackend model and tool usage is metered separately, on a second meter.
Agents API — managed sandboxesmanaged agent runsModel tokens at the selected model's ratesOpenAI tools at standard rates, plus hosted sandboxes at container rates.
Managed agents — tool approvalagent.tool_use eventsToken usage for the turnServer-side evaluation decides run, deny or pause; the decision is auditable per call.

Why it breaks

Why a token cap no longer caps your spend.

There is no route for voice
A gateway that routes Chat Completions, Responses and Realtime has no path for a voice session. Real-time voice needs an explicit endpoint plus WebRTC, WebSocket, sideband or SIP transport — it is not a model swap on an existing route.
A token cap cannot cap a session billed in seconds
The voice session is metered in seconds while the work it delegates is metered in tokens. A spend cap expressed only in tokens governs neither the session 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. Sandbox lifetime is a cost control, not only an infrastructure detail.
The approval boundary is per provider
One provider evaluates tool calls server-side under auto policies and pauses for approval; another leaves permissions, confirmations and function execution to your application. Accountability for a tool call is a per-provider property.
Available today

What ALLM tracks now.

The deployment is the unit of truth. A price, a limit or a lifecycle date belongs to a specific provider endpoint, with the evidence attached.

  • Deployment-level rate cards — input, output and cached-input prices per provider, not per model name.
  • Capabilities and limits — context window, max output, modalities, structured output, tool calling.
  • Lifecycle with qualifiers — release and end-of-life dates, preserving “not before” and “by” as distinct constraints.
  • Evidence on every claim — source URL and observation date attached to each price and lifecycle record.
  • Immutable catalog releases — reproducible snapshots so a routing decision can be audited months later.
Explore the registryRead the docs
In progress

Where this is going.

Cost control that spans meters is the next layer we are building on top of the deployment catalog. It is direction, not a shipped feature — we would rather say so.

  • Non-token meters — voice session seconds, container time and tool invocations normalized beside token rates.
  • Stacked billing — the full cost of a run (model + tools + container), not only its model usage.
  • Meter-aware budgets — spend models that span meters instead of assuming every cost is a token.
  • Voice lifecycle tracking — endpoints that rotate faster than chat endpoints, with their notice periods.

Running agents in production and wrestling with non-token spend? We would like to hear how your budget model is structured — contact@use-allm.com.

Questions

Frequently asked

Why does a token budget no longer cap AI agent spend?

Because the bill now has several meters. A real-time voice session is metered in seconds, an agent sandbox in container time, and tool calls as their own line. A cap expressed only in tokens leaves those meters unmetered, so spend can grow while the token counter looks healthy.

How are real-time voice sessions billed?

OpenAI's GPT-Live 1, released September 10, 2026, is billed at $0.05 per minute, charged per second rather than rounded up to the next minute. Backend model and tool usage triggered by the session is billed separately, so a voice conversation runs on two meters at once. The model is reachable only on v1/live/sessions.

What is billed separately in the OpenAI Agents API?

Agent 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. An agent that spends most of its life waiting in a sandbox carries a cost the token rate does not describe.

Who is accountable for a tool call made by a managed agent?

It depends on the provider. Anthropic's Managed Agents evaluate each agent or MCP tool call server-side under auto permission policies, and the agent.tool_use and agent.mcp_tool_use events report how the call was evaluated — so an approval decision is auditable per call. OpenAI's Agents API leaves permissions, confirmations, private function execution and durable task state with your application.

Are voice endpoints as stable as chat endpoints?

No — they rotate faster. Published lifecycles for voice and realtime endpoints are markedly shorter than for chat models, and some providers have moved to shorter legacy notice periods, including a 45-day option, with access potentially lost after 15 days of inactivity. Business-critical voice traffic belongs on generally available endpoints with the retirement date tracked per provider.

Traceability

Sources

Official provider pages observed September 14, 2026 (Europe/Paris). The live registry renders the deployed records separately.

Start free

Know the cost of a model before you route to it.

Deployment-level pricing, capabilities and lifecycle from official sources — with the evidence attached to every claim.

Get an API key Browse models

Deployment-level rates

Input, output and cached-input prices per provider endpoint.

Sourced and dated

Every price and lifecycle claim carries its evidence.

Snapshot history

Immutable releases for reproducible routing decisions.

Related reading:Voice sessions and agent sandboxes: the non-token meterAll intelligence reports