OpenAI Voice API Pricing: 2026 Prices per Minute and per Call
Published November 20, 2026
Our analysts compiled every voice model price on OpenAI's and Google's API pricing pages on November 20, 2026, traced each price change through 27 archived copies of those pages back to October 2024, and priced the same five-minute inbound lead call on seven OpenAI, Gemini, and Claude setups to show what AI voice costs as of Q4 2026.
Key Findings
- OpenAI's gpt-realtime-2.1 charges $64 per million audio output tokens, about $0.077 for each minute a voice agent speaks. That is 68% less than the $0.24 a minute OpenAI charged when it launched the Realtime API in October 2024.
- GPT-Live-1.1, which OpenAI released on November 19, 2026, bills $0.05 per active minute for both sides of a conversation, plus the backend text model that does the reasoning.
- In our five-minute lead call, gpt-realtime-2.1-mini cost $0.070 and gpt-realtime-2.1 cost $0.23 with the conversation history read from cache. With no cache hits, gpt-realtime-2.1 cost $0.93.
- Gemini 3.8 Live lists $0.005 a minute for caller audio and $0.018 for spoken replies, but our call cost $0.16 because Google re-bills the conversation history at the audio input rate on every turn.
- Anthropic sells no speech model. Paired with OpenAI's streaming transcription and gpt-4o-mini-tts, Claude Haiku 5.5 brought our call to $0.12 and Claude Sonnet 5.5 to $0.13.
Voice agents now answer sales lines, return missed calls, and book appointments for businesses that used to send those calls to voicemail. Most of them run on a speech model from OpenAI or Google, or on a chain of three services: one that turns the caller's speech into text, a text model such as Claude that writes the reply, and one that reads the reply aloud. OpenAI and Google price these models in tokens, a unit that maps poorly onto a phone call, and the two companies count a minute of audio differently. So we converted every price into dollars per minute and per call.
Our analysts read the audio sections of OpenAI's and Google's API pricing pages on November 20, 2026. We converted token prices into per-minute prices with each company's published audio token rates. We dated each price change from OpenAI's changelog and from archived copies of both pricing pages in the Internet Archive. We then built a cost model of a five-minute inbound lead call with 12 back-and-forth turns and priced it at list prices on each setup.
In the following sections, we compare OpenAI's current voice prices per minute, how those prices changed since 2024, what one lead call costs on each setup, Gemini's voice prices, what Claude costs inside a voice agent, transcription prices before OpenAI retires four models in February 2027, and what voice agent platforms charge per minute.
OpenAI Voice API Prices per Minute, November 2026
OpenAI sells models for live voice conversation, text-to-speech, transcription, and live translation, and bills them in several different units. We converted every list price into the cost of one minute of audio so the products can be set side by side.
| Model | What it does | List price | Price per minute |
|---|---|---|---|
| gpt-realtime-2.1 | Speech-to-speech voice agent | $32 in, $64 out per 1M audio tokens | $0.019 caller audio, $0.077 agent speech |
| gpt-realtime-2.1-mini | Lower-cost speech-to-speech | $10 in, $20 out per 1M audio tokens | $0.006 caller audio, $0.024 agent speech |
| gpt-live-1.1 | Full-duplex voice layer with a separate backend model | $0.05 per active minute, $0.02 standby | $0.05 for both sides, plus backend model |
| gpt-realtime-translate | Live speech translation | $0.034 per minute | $0.034 |
| gpt-live-transcribe | Streaming transcription | $0.017 per minute | $0.017 |
| gpt-4o-mini-tts | Text-to-speech | $0.60 per 1M text tokens in, $12 per 1M audio tokens out | $0.015 |
| gpt-transcribe | File and turn transcription | $0.0045 per minute | $0.0045 |

Key Findings
- A minute of agent speech on gpt-realtime-2.1 costs four times as much as a minute of caller audio. OpenAI counts assistant audio at one token per 50 milliseconds and caller audio at one token per 100 milliseconds, and it prices output tokens at twice the input rate.
- GPT-Live-1.1 bills by the clock. Silence before the idle timeout and time spent waiting for backend work count as active time at $0.05 a minute, and the $0.02 standby rate applies only to paused WebRTC sessions.
- On gpt-realtime-2.1, the agent's talk time drives most of the audio bill, so shorter spoken replies lower the cost of every call.
How OpenAI's Realtime Voice Prices Changed Since 2024
OpenAI has cut its realtime audio prices twice since launch and released six new conversational voice models in 2026. We dated each change from OpenAI's changelog and from archived copies of its pricing page.
| Date | Model | Audio price per 1M tokens | Caller audio per minute | Agent speech per minute |
|---|---|---|---|---|
| Oct. 1, 2024 | gpt-4o-realtime-preview | $100 in, $200 out | $0.060 | $0.240 |
| Dec. 17, 2024 | gpt-4o-realtime-preview-2024-12-17 | $40 in, $80 out | $0.024 | $0.096 |
| Dec. 17, 2024 | gpt-4o-mini-realtime-preview | $10 in, $20 out | $0.006 | $0.024 |
| Aug. 28, 2025 | gpt-realtime (general availability) | $32 in, $64 out | $0.019 | $0.077 |
| Oct. 6, 2025 | gpt-realtime-mini | $10 in, $20 out | $0.006 | $0.024 |
| May 7, 2026 | gpt-realtime-2 (text output $24, up from $16) | $32 in, $64 out | $0.019 | $0.077 |
| July 6, 2026 | gpt-realtime-2.1 and gpt-realtime-2.1-mini | $32 and $64; $10 and $20 | $0.019; $0.006 | $0.077; $0.024 |
| Sept. 10, 2026 | gpt-live-1 (gpt-live-1.1 on Nov. 19, 2026) | Billed per minute | $0.05 per active minute, both sides | $0.05 per active minute, both sides |

Key Findings
- The price of a minute of agent speech on OpenAI's top realtime model fell from $0.24 in October 2024 to $0.096 in December 2024 and $0.077 in August 2025. It has not changed through three newer models: gpt-realtime-1.5, gpt-realtime-2, and gpt-realtime-2.1.
- Text output got more expensive. With gpt-realtime-2 in May 2026, OpenAI raised the price of realtime text output tokens, which carry transcripts and reasoning, from $16 to $24 per million, and gpt-realtime-2.1 kept the $24 rate.
- The mini tier has stayed at $10 and $20 per million audio tokens since December 2024, so the flagship's audio output price fell from four times the mini price to 3.2 times.
What a Five-Minute Voice Agent Call Costs on Each Setup
A per-minute price leaves out how often a model re-reads the conversation before each reply. We priced one realistic call, a prospect phoning about a service while the agent qualifies them and offers an appointment, on seven setups at list prices.
| Setup | Cost per call | Cost per minute | Cost of 1,000 calls |
|---|---|---|---|
| gpt-realtime-2.1-mini | $0.070 | $0.014 | $70 |
| Gemini 3.5 Transcribe Live, Claude Sonnet 5.5, and Gemini 3.8 Flash TTS | $0.091 | $0.018 | $91 |
| gpt-live-transcribe, Claude Haiku 5.5, and gpt-4o-mini-tts | $0.116 | $0.023 | $116 |
| gpt-live-transcribe, Claude Sonnet 5.5, and gpt-4o-mini-tts | $0.134 | $0.027 | $134 |
| Gemini 3.8 Live | $0.162 | $0.032 | $162 |
| gpt-realtime-2.1, history cached | $0.233 | $0.047 | $233 |
| gpt-live-1.1 with gpt-6.1-terra as backend | $0.260 | $0.052 | $260 |
| gpt-realtime-2.1, no cache hits | $0.934 | $0.187 | $934 |

Key Findings
- Prompt caching decides the cost of OpenAI's realtime models. With every earlier turn read from cache at $0.40 per million tokens, gpt-realtime-2.1 cost $0.23 for the call. With no cache hits it cost $0.93, four times as much.
- GPT-Live-1.1 came to $0.26 with gpt-6.1-terra as the backend model. The $0.25 session fee made up nearly all of it, and the backend added under one cent.
- At 1,000 calls a month, the setups we priced range from $70 on gpt-realtime-2.1-mini to $260 on GPT-Live-1.1 when caching works, before phone minutes and platform fees.
Gemini Voice API Prices
Google sells live conversation, transcription, translation, and text-to-speech models through the Gemini API, and it prints per-minute equivalents beside most token prices. We list the paid-tier prices, including a text-to-speech increase Google has scheduled for January 1, 2027.
| Model | What it does | List price | Price per minute |
|---|---|---|---|
| Gemini 3.8 Live | Live speech-to-speech | $3 in, $12 out per 1M audio tokens | $0.005 caller audio, $0.018 spoken reply |
| Gemini 2.5 Flash Native Audio | Older live speech-to-speech | $3 in, $12 out per 1M audio tokens | $0.0045 caller audio, $0.018 spoken reply |
| Gemini 3.5 Live Translate | Live speech translation | $3.50 in, $21 out per 1M audio tokens | About $0.037 |
| Gemini 3.5 Transcribe Live | Streaming transcription | $3.50 per 1M audio tokens in, $21 per 1M text tokens out | About $0.009 |
| Gemini 3.5 Transcribe | File transcription | $2 per 1M audio tokens in, $12 per 1M text tokens out | About $0.005 |
| Gemini 3.8 Flash TTS | Text-to-speech | $9 per 1M audio tokens; $18 from Jan. 1, 2027 | $0.0135; $0.027 from Jan. 1, 2027 |
| Gemini 3.8 Flash-Lite TTS | Low-cost text-to-speech | $6 per 1M audio tokens; $12 from Jan. 1, 2027 | $0.009; $0.018 from Jan. 1, 2027 |
| Gemini 2.5 Pro Preview TTS | Text-to-speech | $20 per 1M audio tokens | $0.030 |

Key Findings
- Gemini 3.8 Live charges the same audio rates as Gemini 2.5 Flash Native Audio, $3 and $12 per million tokens, which Google's pricing page has listed for live audio since at least December 2025.
- Google bills the whole session history on every turn of a Live conversation at the audio input rate and lists no cached audio price. Our five-minute call cost $0.16 on Gemini 3.8 Live, against $0.05 if each second of audio were billed once.
- Gemini 3.8 Flash TTS costs $0.0135 a minute through December 31, 2026, less than OpenAI's gpt-4o-mini-tts at $0.015, and its price doubles to $0.027 a minute on January 1, 2027.
What Claude Costs Inside a Voice Agent
Anthropic's API offers text and image models with no speech-to-text, text-to-speech, or speech-to-speech model, so a business that wants Claude on a phone line pairs it with a transcription service and a voice. We priced Claude in our call model and collected the per-minute rates Retell AI, a voice agent platform, charges for Claude, GPT, and Gemini models.
| Model | Maker | Rate per minute |
|---|---|---|
| Claude Haiku 5.5 | Anthropic | $0.0032 |
| Claude Sonnet 5.5 | Anthropic | $0.064 |
| Claude Opus 5.5 | Anthropic | $0.128 |
| GPT 6.1 Terra | OpenAI | $0.032 |
| GPT 6 Sol | OpenAI | $0.064 |
| GPT 6.2 Astra | OpenAI | $0.256 |
| Gemini 3.8 Flash | $0.048 | |
| Gemini 4 Argon | $0.128 |
Key Findings
- In our call model, Claude Haiku 5.5 added $0.001 to a five-minute call and Claude Sonnet 5.5 added $0.019. The transcription and voice services around Claude cost $0.115.
- Retell charges $0.064 a minute for Claude Sonnet 5.5, the same as GPT 6 Sol, and lists both among its three recommended models, next to GPT 6.1 Terra at $0.032.
- Because the speech services cost more than the text model, choosing Claude Sonnet 5.5 over Claude Haiku 5.5 raised our call cost by 1.8 cents.
Transcription Prices and OpenAI's February 2027 Model Retirements
Most voice agents keep a written transcript of each call for the CRM and for quality review. We compared OpenAI's and Google's transcription prices, including four OpenAI models that OpenAI will shut down on February 26, 2027.
| Model | Maker | Price per minute | Status |
|---|---|---|---|
| gpt-transcribe | OpenAI | $0.0045 | Released July 28, 2026 |
| gpt-live-transcribe | OpenAI | $0.017 | Released July 28, 2026, streaming |
| gpt-realtime-whisper | OpenAI | $0.017 | Released May 7, 2026, streaming |
| whisper-1 | OpenAI | $0.006 | Shuts down Feb. 26, 2027 |
| gpt-4o-transcribe | OpenAI | $0.006 | Shuts down Feb. 26, 2027 |
| gpt-4o-mini-transcribe | OpenAI | $0.003 | Shuts down Feb. 26, 2027 |
| Gemini 3.5 Transcribe | About $0.005 | Current | |
| Gemini 3.5 Transcribe Live | About $0.009 | Current, streaming |
Key Findings
- A business on gpt-4o-mini-transcribe at $0.003 a minute will pay 50% more on gpt-transcribe at $0.0045 after February 26, 2027, and nearly six times as much on gpt-live-transcribe at $0.017 if it needs live transcripts.
- Google's streaming transcription costs about $0.009 a minute, about half of OpenAI's $0.017 for gpt-live-transcribe. gpt-4o-transcribe-diarize, which labels each speaker, also shuts down on February 26, 2027.
What Voice Agent Platforms Charge per Minute
Many marketing teams run voice agents through a hosted platform that bundles the models, phone numbers, and call analytics. We collected list rates from Retell AI's pricing page and from a pricing study AssemblyAI published on November 19, 2026.
| Platform and option | Rate | What it covers |
|---|---|---|
| Retell AI, all-in range | $0.07 to $0.31 per minute | Voice agent minutes, depending on the model and voice chosen |
| Retell AI, GPT Realtime 2.1 | $0.38 per minute | OpenAI speech-to-speech model |
| Retell AI, GPT Realtime 2.1 mini | $0.07 per minute | OpenAI lower-cost speech-to-speech model |
| Retell AI, GPT-Live 1 | $0.05 per minute plus backend model | OpenAI voice layer |
| Retell AI, voice infrastructure | $0.055 per minute | Platform fee on chained setups, before model and voice |
| Vapi | $0.05 per minute | Hosting, with models billed at provider cost |
| ElevenLabs Agents | $0.08 per minute beyond plan minutes | Agent minutes, with model usage billed separately |
Key Findings
- Retell charges $0.38 a minute for GPT Realtime 2.1, or $1.90 for a five-minute call, about eight times the $0.23 we calculated at OpenAI's list prices with caching.
- At Retell's rates, a five-minute call with Claude Sonnet 5.5, Retell's voice infrastructure, and a Retell platform voice costs $0.67 before phone minutes, which Retell lists at $0.015 a minute in the US.
- AssemblyAI's study put the same five-minute call at $0.35 to $1.55 on Retell, $0.40 on ElevenLabs Agents, and $0.25 plus model costs on Vapi.
What Voice API Prices Mean for AI Phone Agents
For a business answering inbound leads with an AI agent, the model cost of a five-minute call at list prices runs from about 7 cents to 26 cents when caching works. Platform fees and phone minutes can add several times that amount, as Retell's $0.38 a minute for GPT Realtime 2.1 shows.
OpenAI's voice pricing has moved toward the clock. GPT-Live-1 and GPT-Live-1.1 charge by the minute, OpenAI's newest transcription and translation models charge by the minute, and Google prints per-minute equivalents beside its token prices.
Google's text-to-speech increase takes effect first. On January 1, 2027, Gemini 3.8 Flash TTS goes from $9 to $18 per million audio tokens, and eight weeks later, on February 26, 2027, OpenAI shuts down whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, and gpt-4o-transcribe-diarize.
Data: openai-voice-api-pricing-2026.csv
Data Collection and Methodology
Source 1 is our own research, compiled on November 20, 2026: a table of every voice model price on OpenAI's and Google's API pricing pages, a dated record of OpenAI's realtime price changes, and a cost model of one inbound lead call.
Prices are paid-tier list prices in US dollars, read from the OpenAI and Google pricing pages on November 20, 2026. Price history comes from OpenAI's API changelog and from 27 archived copies of the OpenAI and Gemini pricing pages in the Internet Archive, captured from October 2024 to November 2026. OpenAI counts caller audio at one token per 100 milliseconds (600 tokens a minute) and assistant audio at one token per 50 milliseconds (1,200 a minute), per its Realtime API cost guide. Google counts audio at 25 tokens a second (1,500 a minute). Where a company publishes its own per-minute estimate, we use it.
The call model is a five-minute inbound call with 12 turns: the caller speaks for two minutes, the agent speaks for two minutes, and one minute is pauses. Instructions and tool definitions take 1,500 text tokens, and spoken words run about 200 text tokens a minute. On OpenAI's realtime models, each turn re-sends the instructions and the earlier conversation, and the agent's audio carries text tokens equal to one third of its audio tokens, the ratio in OpenAI's own usage example. The cached case reads all earlier turns at the cached input rate; the no-cache case pays the full input rate. On Gemini 3.8 Live, each turn re-bills the earlier conversation's audio at the audio input rate, as Google's Live API documentation describes, and fresh input covers the three minutes the model is listening.
GPT-Live-1.1 is billed at $0.05 for each of the five minutes, with gpt-6.1-terra called on every turn as the backend model. The chained setups bill streaming transcription for the full five minutes, text-to-speech for the agent's two minutes of speech, and the text model per turn, with the instructions cached after the first turn. Reasoning tokens, tool calls, telephony, and platform fees are left out. Figures from Retell AI and AssemblyAI are reported as each company published them.
Sources
- Top GEO Agencies, OpenAI and Gemini voice API price table, price history, and five-minute call cost model (Source 1, our research), compiled Nov. 20, 2026. Download the data
- OpenAI, API pricing, accessed Nov. 20, 2026. platform.openai.com
- OpenAI, Voice cost optimization and Realtime API costs, accessed Nov. 20, 2026. developers.openai.com
- OpenAI, API changelog, accessed Nov. 20, 2026. developers.openai.com
- OpenAI, Build more natural voice experiences with GPT-Live-1 in the API, Sept. 10, 2026. openai.com
- OpenAI, GPT-Live-1 model page, accessed Nov. 20, 2026. developers.openai.com
- OpenAI, Advancing voice intelligence with new models in the API, May 7, 2026. openai.com
- OpenAI, GPT-Realtime-2 model page, accessed Nov. 20, 2026. developers.openai.com
- Internet Archive, OpenAI API pricing page captures, Dec. 5, 2024, Jan. 9, 2025, and Sept. 2, 2025. web.archive.org
- Internet Archive, OpenAI API docs pricing page captures, Feb. 10 to Nov. 1, 2026. web.archive.org
- Google, Gemini Developer API pricing, accessed Nov. 20, 2026. ai.google.dev
- Internet Archive, Gemini API pricing page captures, June 3, 2025 to Nov. 1, 2026. web.archive.org
- Google Cloud, Start and manage live sessions, accessed Nov. 20, 2026. docs.cloud.google.com
- Duplex Compute, Gemini Live API pricing: why a long conversation costs more per minute, Oct. 13, 2026. duplexcompute.com
- Anthropic, Claude API pricing, accessed Nov. 20, 2026. platform.claude.com
- Anthropic, Claude API overview, accessed Nov. 20, 2026. platform.claude.com
- Retell AI, AI Phone Agent Pricing, accessed Nov. 20, 2026. retellai.com
- AssemblyAI, AI voice agent pricing in 2026: per-minute platforms vs stacked per-layer costs, Nov. 19, 2026. assemblyai.com
- Microsoft, A New Chapter for Realtime AI: Reasoning, Translation, and Real-Time Transcription, May 7, 2026, updated July 8, 2026. techcommunity.microsoft.com
- MarkTechPost, OpenAI Releases GPT-Realtime-2.1 and GPT-Realtime-2.1-mini, July 6, 2026. marktechpost.com
Source: OpenAI API pricing page and Realtime API cost guide, read Nov. 20, 2026; per-minute conversions by Top GEO Agencies.
Source: OpenAI API changelog; OpenAI pricing page captures in the Internet Archive from Dec. 5, 2024, Jan. 9, 2025, Sept. 2, 2025, Dec. 3, 2025, and Feb. 10 to Nov. 1, 2026; OpenAI, Sept. 10, 2026.
Source: Top GEO Agencies cost model at OpenAI, Google, and Anthropic list prices, Nov. 20, 2026. Excludes telephony and platform fees.
Source: Google, Gemini Developer API pricing page, read Nov. 20, 2026; Google per-minute estimates where given, otherwise converted at 25 audio tokens per second.
Source: Retell AI pricing page, read Nov. 20, 2026; Anthropic, Claude API pricing and API overview, read Nov. 20, 2026.
Source: OpenAI API pricing page and changelog (Aug. 26, 2026 deprecation notice); Google, Gemini Developer API pricing page; read Nov. 20, 2026.
Source: Retell AI pricing page, read Nov. 20, 2026; AssemblyAI, AI voice agent pricing in 2026, Nov. 19, 2026 (Vapi and ElevenLabs rates).