GPT-Live-1 Cost Calculator: What a Voice Agent Really Costs Per Call

The short answer: GPT-Live-1 is billed on two meters, and the $0.05 per minute you have read about is only the first one. Voice time is $0.05 per minute, billed by the second, with no round-up — but the agent's backend reasoning, tool calls and web search are billed separately at the backend model's normal token rates. On the default case on this page (4-minute calls, 1,000 calls a day, 6 reasoning turns per call on GPT-5.6 Terra) voice alone is $200.00 per day and the backend adds $39.60 per day on top. Change the backend from Terra to GPT-6 Astra and the backend line becomes $180.00 per day — within 10% of the voice bill, for a model that is doing the thinking, not the talking.

That failure mode is the reason this calculator exists. Nearly every published breakdown of the GPT-Live-1 API stops at the $0.05 per minute figure, which is a voice number. A production phone agent is a voice session plus a reasoning workload, and the second half scales with how hard the caller's request is, not with how long the call lasts. It also has a capacity dimension nobody models: GPT-Live session limits are measured in concurrent sessions, not dollars, so a deployment can run out of session headroom before it runs out of budget.

Everything below is computed live from editable inputs. Published OpenAI rates are labelled as such; every modelling assumption (call mix, token shapes, peak factor, month definition) is labelled as ours and is editable. The machine-readable export at the bottom of the calculator is the exact parameter set and result set the article and this page cite.

—

How the two meters work

Meter one is duration. A GPT-Live-1 session is billed on active session time at $0.05 per minute. OpenAI's cost guide defines active session time as including periods when the user speaks, the assistant speaks, both are silent, and the backend is working — and it notes that muting the microphone does not close the session. Two consequences follow: silence is billable, and delegating a slow reasoning task to the backend does not pause the voice meter. Billing is by the second with no round-up, so a 40-second call is $0.0333 rather than a full minute.

Meter two is reasoning. Backend Responses calls use the normal pricing for the configured model and tools, and OpenAI's cost guide says to include model input and output tokens, cached input where supported, and any applicable image or tool charges — plus the cost of any other services your application calls. The published WebRTC example delegates to GPT-5.6 Terra with hosted web search, so the default on this page uses Terra at its list rates: $2.00 per 1M input tokens, $0.20 cached input, $12.00 per 1M output tokens (short-context column).

The 15-second session-init rule, precisely. Creating a WebRTC session with POST /v1/live/sessions bills 15 seconds of voice duration while the session initializes, and that amount is credited against duration charges once the session starts running. OpenAI's own worked example bills a 90-second session as 90 seconds, "not 105 seconds". The net effect is therefore a 15-second floor, max(session_seconds, 15), not a 15-second adder — and both of the primary statements scope it to WebRTC session creation. The WebSocket guide documents no init charge, which is why transport is a toggle here.

Usage accounting. For reconciliation, session.usage.updated replaces a cumulative duration snapshot rather than adding to it; record the final usage.seconds from session.closed once. Backend work can also outlive the voice session, so a deployment can accrue backend cost with zero voice cost and the reverse.

Why the backend line decides whether this product works

The voice meter is linear, predictable and cheap: 4 minutes is 20 cents, no matter what the caller asks. The backend meter is neither. It scales with turns, with context, with how many turns trigger a tool or a search, and with the price of the model you delegated to. On this page's defaults the backend is a modest $0.0396 per call against $0.2000 of voice time. Switch to Astra, OpenAI's suggested model for complex customer issues, and the same token shape costs $0.1800 per call — roughly 5× the Terra backend, or 90% of the entire voice bill.

That is the "cheap voice, expensive reasoning" failure mode in one number. A deployment priced on the $0.05 per minute rate, then handed to an expensive reasoner with a high tool-call share, can double or triple its unit economics without a single line of its voice code changing. The calculator's backend-share line exists to make that visible before it shows up on an invoice.

Derived from OpenAI's published per-1M token rates (Standard, short-context column) applied to the same token shape; reproduce any row in the calculator above. Cache-write tokens are not modelled because they are paid once per cached prefix rather than per turn.

Sizing the line: concurrency, not dollars

GPT-Live rate limits are measured in concurrent sessions, and the ceilings are 25 / 50 / 200 / 300 / 500 for Tiers 1–5. The Free tier is not supported. That makes capacity a first-class cost question: a deployment can be comfortably affordable and still fail at 9 a.m. because too many callers are talking at once.

The sizing formula is Little's Law applied to a call centre line:

At 1,000 calls a day in an 8-hour window with 4-minute calls, that is 125 calls per hour × 240 seconds ÷ 3600 = 8.33 concurrent sessions on average. A 2.5× busy-hour peak takes it to 20.8 — still inside Tier 1's ceiling of 25, with roughly 17% headroom left. Push the same call volume into a 4-hour window, or let average calls run to 6 minutes, and the peak requirement crosses 25 and you are capacity-limited before you are budget-limited: the constraint is session headroom, not spend, and the fix is a tier change rather than a cheaper model.

It is worth noticing how far apart the two constraints sit. Tier 1 qualification starts at $5 paid and $100 per month of API spend, and the default case on this page burns $200.00 of voice time per day. Spend is not what limits a voice deployment of this size; concurrent sessions are. That is exactly why the planner is on the page.

What is not included in the $0.05 per minute

• Backend model tokens — input, cached input and output tokens at the delegated model's rates.
• Tool calls and web search — billed at model rates, and search results injected into context are input tokens you pay for.
• Any other service in your call path — telephony, transcription you run yourself, your own retrieval stack. OpenAI's cost guide says to include them in your estimate.
• Idle voice time — already inside the $0.05, and easy to forget: it is roughly $0.05 per idle minute, and closing the session is what stops it.
• Custom voices — sales-gated with no published rate, so they are deliberately not modelled.

Methodology and provenance

Every number on this page is either (a) an OpenAI-published rate reproduced from the pages listed under Sources, (b) arithmetic derived from those rates, or (c) a modelling assumption labelled as such in the calculator and in the JSON export. Default parameters and their provenance: voice rate, init seconds and scope, tier ceilings and backend rates are OpenAI-published; calls per day (1,000), call length (4:00), operating window (8 h), peak factor (1×), turns per call (6), tokens per turn (1,500 in / 300 out), tool-call share (35%), cached-input share (0%), idle share (30%) and the 30-day month are our modelling choices, chosen to match a mid-size phone-support deployment and fully editable.

The month convention is printed on every monthly row because OpenAI defines no month. This page's default is a 30-day month ($200.00 per day = $6,000.00); the same day rate is $6,200.00 over 31 days and $6,083.33 at 365÷12, and both alternatives are shown next to the primary figure.

Known limits of this model. The published short-context / long-context rate boundary is not stated by OpenAI, so the long-context toggle is manual. Cache-write tokens are not modelled. The tokens-per-minute rates in the Realtime comparison are ours, because OpenAI publishes no conversion from audio minutes to audio tokens — that is why the comparison prints a break-even token rate as well as a cost. Tool and search spend is a per-invocation input, not an estimate of search-result token volume. And no benchmark figure on this page is presented as independently reproduced: OpenAI's published evaluations are OpenAI's.

Frequently asked questions

Sources

• OpenAI — Introducing GPT-Live-1 in the API (announcement, September 10, 2026): $0.05 per minute, backend delegation, rate-limit framing.
• GPT-Live-1 model card: model ID, supported endpoints, concurrent-session rate limits, Free tier not supported.
• OpenAI API pricing: GPT-Live session rate, per-1M backend token rates (short and long context), data-residency uplift, GPT-Realtime-2.1 rows.
• Voice latency and cost guide: the 15-second WebRTC initialization charge and its credit rule; the worked 90-second example; the billable-time definition.
• Live sessions guide and conversations guide: duration billing, transports, session lifecycle and usage events.
• Delegation guide and WebRTC guide: backend model choice, service tiers, the GPT-5.6 Terra example.
• WebSocket guide and rate-limit guide: the server-side transport and the tier qualification thresholds.