Direct answer: you cannot compare realtime voice agents on "price per million tokens" alone, because vendors slice audio into tokens at different rates. OpenAI bills user audio at 1 token per 100 ms (600 tokens/minute) and assistant audio at 1 token per 50 ms (1,200 tokens/minute); Google Live uses 25 tokens/second (1,500 tokens/minute). Converted at full-duplex usage, Google Gemini 3.8 Live is about USD 0.0225/minute, OpenAI realtime mini about 0.030, GPT-Live a flat 0.05, and OpenAI realtime 2.1 about 0.096. A 3-minute call therefore costs roughly USD 0.068–0.288 — but longer calls with more turns cost more than minutes × rate.
This guide is for product, customer-service and procurement teams building phone agents, voice assistants or live translation. Data cut-off: 1 November 2026; prices come from official OpenAI and Google Cloud pages, in USD, excluding tax, with no FX conversion. The worked examples are editorial models, not quotes; telephony, numbers and transcription add-ons must be priced with the actual provider.
Convert token rates to cost per minute before comparing unit prices; token-billed plans are cheaper when calls are short and silent, while per-second/minute plans are easier to forecast on long calls; resent conversation history and transcription add-ons are the two main reasons invoices exceed minutes × rate.
Why can't you compare "per million tokens"?
Because audio is tokenised at different rates: OpenAI input at 10 tokens/second and output at 20 tokens/second; Google Live at 25 tokens/second. One minute of audio can be billed as 600 tokens or 1,500 tokens depending on vendor and direction.
| Vendor / model | Official audio token rate | Tokens per minute | Official price (per 1M tokens) |
|---|---|---|---|
| OpenAI realtime (user audio) | 1 token per 100 ms | 600 | Input USD 32 (2.1) / 10 (mini) |
| OpenAI realtime (assistant audio) | 1 token per 50 ms | 1,200 | Output USD 64 (2.1) / 20 (mini) |
| Google Gemini 3.8 Live | 25 tokens/second | 1,500 | Input USD 3 / output USD 12 (non-global) |
| OpenAI GPT-Live | Not token-based; billed per second | — | USD 0.05/minute (per second, no rounding up) |
Sources: OpenAI "Realtime API costs" (audio token rules), OpenAI API pricing and Google Cloud Vertex AI pricing (Gemini 3.8 Live audio and Live API billing notes). Public list prices, excluding tax and add-ons.
What is the cost per minute?
At full-duplex usage: Gemini 3.8 Live about USD 0.0225/minute, OpenAI realtime mini about 0.030, GPT-Live 0.05, and OpenAI realtime 2.1 about 0.096; add live transcription at about 0.017/minute.
| Option | Calculation | USD/minute | 3-minute call |
|---|---|---|---|
| Gemini 3.8 Live | 1,500 × 3 / 1M + 1,500 × 12 / 1M | 0.0225 | 0.0675 |
| OpenAI realtime 2.1 mini | 600 × 10 / 1M + 1,200 × 20 / 1M | 0.030 | 0.090 |
| OpenAI GPT-Live 1 | Flat per minute | 0.050 | 0.150 |
| OpenAI realtime 2.1 | 600 × 32 / 1M + 1,200 × 64 / 1M | 0.096 | 0.288 |
| Add-on: live transcription | gpt-live-transcribe at 0.017/min | +0.017 | +0.051 |
Editorial model: assumes the user and assistant each speak for the full 60 seconds; real calls contain silence, interruptions and waiting, so token-billed options will come in lower. Excludes phone numbers, telephony, storage and tax.
What does 1,000 calls cost?
At 3 minutes per call (3,000 minutes): Gemini USD 67.50, mini 90, GPT-Live 150, realtime 2.1 288; enabling live transcription adds about USD 51.
| Option | 3,000 minutes | With live transcription |
|---|---|---|
| Gemini 3.8 Live | USD 67.50 | USD 118.50 |
| OpenAI realtime 2.1 mini | USD 90.00 | USD 141.00 |
| OpenAI GPT-Live 1 | USD 150.00 | USD 201.00 |
| OpenAI realtime 2.1 | USD 288.00 | USD 339.00 |
Editorial model: 3 minutes × 1,000 calls = 3,000 minutes; live transcription at USD 0.017/min × 3,000 = 51. Excludes telephony, numbers, concurrency and retries; the platform invoice governs.
Why does the invoice exceed minutes × rate?
Because conversation history is billed again: OpenAI resends the entire conversation for every response, and Google bills the session context window turn by turn. More turns mean larger input. Caching and truncation help but do not eliminate the growth.
| Factor | Official rule | Invoice impact |
|---|---|---|
| Resent history | OpenAI: the whole conversation is sent for each response | Turn 10 costs much more than turn 1 |
| Context-window billing | Google: billed per turn including all prior tokens | Long sessions grow roughly quadratically |
| Caching | OpenAI: automatic prompt caching, cached portion at cached rate | Can cut repeated input sharply |
| Truncation | OpenAI: configurable token limit and retention ratio | Lower input per turn, may lose memory |
| Silence | GPT-Live: active session includes silence and backend waiting | Per-second billing charges idle time |
| Initialisation | GPT-Live: WebRTC session creation bills 15 seconds upfront | Frequent reconnects add up |
| Transcription/translation | OpenAI: billed separately per minute | Adds 0.017–0.034/minute |
Official rules from OpenAI's Realtime cost documentation and Google Cloud's Live API pricing notes; cached rates and truncation behaviour can differ by model and region.
How can you lower the per-minute cost?
Five moves: test the mini model, enable caching, set truncation, pre-load known context, and close sessions when the task ends; the first four cut tokens, the last cuts duration.
| Lever | How | Expected impact |
|---|---|---|
| Start with mini | Tune prompts on the larger model, then validate on mini | Unit price drops to roughly one third |
| Prompt caching | Keep instructions and tool definitions stable at the front | Lower repeated input cost |
| Truncation | Cap input tokens per turn and retention ratio | Controls long-session growth |
| Pre-load context | Put order number, status and known facts into the session | Fewer back-and-forth turns |
| Close promptly | Close when the task completes; do not leave sessions idle | Direct cut in per-second/minute fees |
| Lean tool calls | Parallel lookups, less serial waiting | Shorter sessions |
Editorial guidance; the "one third" estimate uses the published mini versus 2.1 audio prices and varies with the input/output mix.
Per minute or per token — which should you choose?
Short, query-style calls with silence favour token billing; long calls needing predictable budgets favour per-second/minute billing; when transcription or translation is involved, include the per-minute add-ons.
| Scenario | Prefer | Why | Do not ignore |
|---|---|---|---|
| Short Q&A / lookups | Per token | Silence and short replies are not billed | Turn growth |
| Long calls / agent assist | Per second or minute | Predictable budget | Silence is billed |
| Live translation | Per minute | Duration-based pricing | Two-way audio |
| Transcribed QA | Tokens + transcription | Transcription is separate | Storage and retention |
| High-volume outbound | Per minute + concurrency cap | Capacity can be sized | Retries and failures |
Editorial guidance, not a vendor ranking; run 50–100 real calls before launch and calibrate against the invoice.
What should procurement and acceptance check?
Token rates, billing unit, history resend and cache rules, truncation support, transcription add-ons, concurrency and reconnect billing — not just the per-million-token price.
| Acceptance item | Suggested definition | Evidence to keep |
|---|---|---|
| Billing unit | Token / second / minute, and rounding rules | Price screenshots, sample invoice |
| Token rate | Input and output converted separately to per minute | Official documentation |
| History and cache | Whether history is resent; cached-input price | Usage events, invoice lines |
| Truncation | Can input per turn and retention be capped | Config, regression results |
| Add-ons | Transcription, translation, storage prices | Price list, invoice |
| Concurrency | Limits and initialisation billing | Load test, contract |
| Cost per call | Total cost ÷ successfully completed calls | Call logs, invoice |
| Failures | Whether failed calls are billed | Invoice, retry log |
Editorial guidance, not legal advice; recording, transcription and personal-data retention need separate assessment under applicable law.
Who should read this — and who should not?
For teams building phone agents, voice assistants or live translation. Not for text-only support, or projects without real call-duration and turn data.
Read if: you work in customer service, product, engineering or procurement; you are comparing OpenAI Realtime, GPT-Live and Gemini Live; you need a budget for 1,000 calls. Skip if: you only run text chat; you have no real call samples; you want this article to pick a vendor. Supports: per-minute conversion, billing-model choice, optimisation list and acceptance criteria.
Lingyi does not rank voice models or promise call quality. To connect voice, transcription, knowledge base and ticketing, see LingHarness; to fold voice cost into project budgets, talk to us.
Related reading: AI customer-service cost and resolution, speech-to-text cost, and AI voice-over cost; this article only covers realtime voice token rates, per-minute cost and billing models.
FAQ
How is realtime voice billed?
Three ways: per token (OpenAI Realtime slices audio at 100 ms / 50 ms), per second or minute (OpenAI GPT-Live at USD 0.05/minute, per second without rounding up), and per minute (translation and transcription, e.g. gpt-live-transcribe at USD 0.017/minute). Convert to per minute before comparing.
What does one minute cost?
At full-duplex usage: Gemini 3.8 Live about USD 0.0225, OpenAI realtime mini about 0.030, GPT-Live 0.05, and OpenAI realtime 2.1 about 0.096. Real calls contain silence, so token-billed options usually come in below the full-duplex figure.
Why can't we compare per-million-token prices?
Because token rates differ: OpenAI input is 10 tokens/second and output 20 tokens/second, while Google Live is 25 tokens/second. One minute is 1,500 tokens on Google but only 600 input tokens on OpenAI; price and token count must be combined.
Why do longer calls cost more?
Every response sends the conversation history again (OpenAI states the whole conversation is resent; Google bills the session context window turn by turn), so input tokens grow with turns. Caching and truncation reduce but do not remove the growth.
Does GPT-Live bill silence?
Yes. Active session time includes user speech, assistant speech, mutual silence and backend processing; WebRTC session creation also bills 15 seconds upfront (credited later). Close sessions promptly when the task is done.
How much does live transcription add?
It is billed per minute: gpt-live-transcribe at USD 0.017/minute and gpt-realtime-translate at USD 0.034/minute. At 1,000 calls × 3 minutes (3,000 minutes), transcription adds about USD 51.
How should we estimate our own bill?
Run 50–100 real calls, record duration, turns, input/output tokens and failure rate, then apply the rates; do not rely on minutes × rate alone, and do not use a vendor demo's call length as your own.
Method and limitations
This article uses only official pages accessible on 1 November 2026: OpenAI's "Realtime API costs" and API pricing, and Google Cloud's Vertex AI pricing (Gemini 3.8 Live and Live API billing notes). All amounts are converted from public list prices and official token rates; no enterprise agreements or reseller quotes are used. USD is not converted to RMB and no vendor ranking is implied. Token rates, regional prices, caching and truncation rules can change, and Google's 3.8 Live price is marked non-global. Per-minute costs assume both parties speak for the full 60 seconds; real silence, interruptions and turns change the result. The 3-minute call and 1,000-call figures are example assumptions, not benchmarks. The article does not assess voice quality, latency or recognition accuracy and is not legal advice.
Sources and basis
Data cut-off: 1 November 2026. The following are official pages; prices, rates and terms follow the original text and invoice.
- OpenAI, Realtime API costs: platform.openai.com/docs/guides/voice-latency-cost. User audio 1 token per 100 ms, assistant audio 1 token per 50 ms; whole conversation resent per response; caching, truncation, GPT-Live per-second billing and the 15-second WebRTC initialisation.
- OpenAI, API pricing: platform.openai.com/docs/pricing. gpt-realtime-2.1 audio at USD 32/64, mini at 10/20 per 1M tokens; gpt-live-1 USD 0.05/minute; gpt-live-transcribe USD 0.017 and realtime-translate USD 0.034 per minute.
- Google Cloud, Vertex AI generative AI pricing: cloud.google.com/vertex-ai/generative-ai/pricing. Gemini 3.8 Live input audio USD 3 and output audio USD 12 per 1M tokens (non-global); Live API audio at 25 tokens/second; per-turn session context-window billing.