OpenAI Realtime API Pricing (2026): Cost per Minute
OpenAI doesn't sell its Realtime API by the minute. You pay per token for the audio and text going into and out of the model each time it replies: gpt-realtime-2.1 is $32 per 1M audio tokens in ($0.40 cached) and $64 out, and gpt-realtime-2.1-mini is $10 in ($0.30 cached) and $20 out. There's no platform fee, no monthly plan and no charge for the connection, but OpenAI doesn't sell phone numbers, so your SIP carrier bills the line. Every price below is from OpenAI's API pricing page and its Realtime cost guide (developers.openai.com), read October 25, 2026. The per-minute figures are Cozzo's arithmetic from OpenAI's token rules, not a bill we measured.
OpenAI bills tokens, not minutes, and every reply re-sends the whole call so far. Keep your instructions and tools the same for the whole call so that history hits the cache, and keep the agent's answers short. Agent speech costs four times what caller speech does.
gpt-realtime-2.1 costs $32/1M audio tokens in, $64 out. What that means per minute, per call and per month, why long calls cost more, and what mini saves.
What you pay for on an OpenAI Realtime call
Caller audio (input)
$0.006–$0.0192 per min of caller speech
OpenAI counts caller audio at 1 token per 100 ms, so 600 tokens a minute: $0.0192 at gpt-realtime-2.1's $32 per 1M, or $0.006 on mini at $10. Voice activity detection filters out silence, so empty audio isn't billed.
Agent speech (output)
$0.024–$0.0768 per min of agent speech
Agent audio counts at 1 token per 50 ms, so 1,200 tokens a minute: $0.0768 at $64 per 1M on gpt-realtime-2.1, or $0.024 on mini at $20. Twice the tokens at twice the rate, so a minute of agent talking costs 4× a minute of caller talking.
The conversation so far (re-sent)
Grows with every turn
Each reply sends the whole conversation back in as input, including the agent's own earlier audio. Cached input is $0.40 per 1M on gpt-realtime-2.1 ($0.30 on mini); anything not cached is billed at the full $32 ($10). OpenAI says caching is automatic but best-effort, not guaranteed.
Text tokens
$4 in / $24 out per 1M
Your instructions, tool definitions, tool results and any text replies. gpt-realtime-2.1 text is $4 in ($0.40 cached) and $24 out per 1M tokens; mini is $0.60 in ($0.06 cached) and $2.40 out. Instructions and tools are re-sent every turn too.
Caller transcripts (optional)
$0.0045/min if turned on
Input transcription is billed separately. OpenAI's cost guide recommends gpt-transcribe for it, listed at $0.0045 a minute. Leave it off if you don't need text transcripts of the caller.
Phone line (your SIP carrier)
Your carrier's rate
OpenAI says there's no charge for bandwidth or connections, and its SIP endpoint takes calls from your own trunk (Twilio, Telnyx, Bandwidth and others). On Twilio a US local number is $1.15 a month and inbound calls $0.0085 a minute.
Total per minute
about 5¢–21¢ a minute on gpt-realtime-2.1
Cozzo's worked example below (a 3-minute, 12-turn call with equal talk time), audio only. The low end assumes the re-sent history is cached; the high end assumes none of it is. On mini the same call is about 1.6¢–6.5¢ a minute. Text tokens and your carrier come on top.
What a month costs on OpenAI's own rates
500 min/month
- 2.1, history cached
- $25
- 2.1, no cache hits
- $103
- mini, history cached
- $8
- mini, no cache hits
- $32
1,000 min/month
- 2.1, history cached
- $50
- 2.1, no cache hits
- $206
- mini, history cached
- $17
- mini, no cache hits
- $65
2,500 min/month
- 2.1, history cached
- $125
- 2.1, no cache hits
- $516
- mini, history cached
- $41
- mini, no cache hits
- $161
5,000 min/month
- 2.1, history cached
- $250
- 2.1, no cache hits
- $1,032
- mini, history cached
- $83
- mini, no cache hits
- $323
10,000 min/month
- 2.1, history cached
- $500
- 2.1, no cache hits
- $2,064
- mini, history cached
- $165
- mini, no cache hits
- $645
Cozzo's arithmetic from OpenAI's published rates and token rules, using the worked example's per-minute cost: $0.0500 or $0.2064 a minute on gpt-realtime-2.1 and $0.0165 or $0.0645 on mini, depending on whether the re-sent conversation hits the cache. Audio tokens only: no text tokens for instructions and tools, no transcripts, and no carrier minutes or numbers. Real calls with long prompts or long turns will cost more. Want your own volume and carrier? Try the cost calculator.
Packages and add-ons
Cozzo's worked example: a 3-minute call
$0.15–$0.62 per call
12 back-and-forth turns, caller and agent each speaking 7.5 s per turn, on gpt-realtime-2.1. New caller audio is 900 tokens ($0.029), agent speech 1,800 tokens ($0.115), and the re-sent conversation adds up to 14,850 audio tokens: $0.006 if it's cached, $0.475 if it isn't. That's about 5¢ or 21¢ a minute. On mini the same call is $0.05–$0.19.
gpt-realtime-2.1-mini
$10 in / $20 out per 1M audio
About a third of gpt-realtime-2.1's audio prices, with cached audio at $0.30 per 1M and text at $0.60 in / $2.40 out. OpenAI says the trade-off is instruction following and function calling, and recommends getting your agent working on the full model first, then trying mini.
GPT-Live 1 voice sessions
$0.05/min, billed per second
OpenAI's other voice option prices the voice session by time instead of tokens: $0.05 a minute, billed per second, counting silence and the time the backend is working. The backend model and tools are billed separately at their own rates. Creating a WebRTC session bills 15 seconds up front, which is credited once the session runs.
Older realtime models
Same audio price
gpt-realtime-2, gpt-realtime-1.5 and gpt-realtime are still listed at $32 in / $64 out per 1M audio tokens, and gpt-realtime-mini at $10 / $20. Text output is $16 per 1M on gpt-realtime-1.5 and gpt-realtime against $24 on the 2.x models.
Data residency
+10%
OpenAI's pricing page says regional processing (data residency) endpoints cost 10% more for models released on or after March 5, 2026, and FedRAMP endpoints also carry a 10% uplift.
No plan, no minimum
Pay per token
There's no monthly fee, platform fee or per-minute floor on the Realtime API itself. Costs are charged as each reply is created, and you can read the exact token counts for every reply in the response.done event or the Realtime Playground's Logs tab.
How it compares on published pricing
OpenAI Realtime API- Starting price
- gpt-realtime-2.1-mini: audio $10 in / $20 out per 1M tokens
- Best for
- Developers building speech-to-speech voice agents directly on OpenAI models
Gemini Live API- Starting price
- Free tier; paid audio ~$0.005/min in, ~$0.018/min out (Gemini 3.8 Live)
- Best for
- Developers who want a realtime speech-to-speech model from Google
Ultravox- Starting price
- $0.05/min (first 30 min free)
- Best for
- Developers wanting a speech-native model at a flat per-minute rate
Vapi- Starting price
- $0 to start; $0.05/min Vapi hosting, models billed pass-through at cost
- Best for
- Developers who want to choose their own LLM, transcriber, voice and telephony provider
Retell AI- Starting price
- $0 platform fee; $0.07–$0.31/min for voice agents
- Best for
- Teams that want a no-code flow builder plus an API with itemized per-minute pricing
ElevenLabs Agents- Starting price
- Agents: $0.08 per minute
- Best for
- Builders who want ElevenLabs' voices and speech models in one agent pipeline
LiveKit Agents- Starting price
- Free; Ship from $50/mo
- Best for
- Developers building on open-source realtime infrastructure
Pricing explained, line by line:Ultravox pricing →Vapi pricing →Retell AI pricing →ElevenLabs Agents pricing →
Cozzo-tested · Round 4 · October 22, 2026
AI receptionist benchmark: 8 voice agents on 5 real-call scenarios
OpenAI's Realtime API (gpt-realtime-2.1 over OpenAI's SIP endpoint) passed 12 of 15 Round 4 calls with a median reply of 1.71 seconds. It passed none of the three new-patient booking calls, which need a listed slot plus a correct read-back of the caller's name and number. Here's how every platform did on the same dental-office calls. Each chip shows calls passed out of 3. Fastest median reply first.
ElevenLabs · 13/15 · 1.06 s median reply
- Book a new patient
- Reschedule
- Unknown insurance
- Interruption
- Dental emergency
Ultravox · 14/15 · 1.27 s median reply
- Book a new patient
- Reschedule
- Unknown insurance
- Interruption
- Dental emergency
Millis AI · 12/15 · 1.31 s median reply
- Book a new patient
- Reschedule
- Unknown insurance
- Interruption
- Dental emergency
Retell AI · 12/15 · 1.38 s median reply
- Book a new patient
- Reschedule
- Unknown insurance
- Interruption
- Dental emergency
OpenAI Realtime · 12/15 · 1.71 s median reply
- Book a new patient
- Reschedule
- Unknown insurance
- Interruption
- Dental emergency
Vapi · 9/15 · 1.90 s‡ median reply
- Book a new patient
- Reschedule
- Unknown insurance
- Interruption
- Dental emergency
Bland AI · 10/15 · 1.94 s median reply
- Book a new patient
- Reschedule
- Unknown insurance
- Interruption
- Dental emergency
LiveKit · 11/15 · 2.00 s median reply
- Book a new patient
- Reschedule
- Unknown insurance
- Interruption
- Dental emergency
‡ Vapi’s 1.90 s median is end-to-end on an untuned default setup, not comparable to Vapi’s published figures, which leave out endpointing and network time. † Callback taken, but the number was read back wrong at least once. * Reply heard each time, but its content wasn’t checked against the vendor’s own transcript. 3 calls per scenario, one prompt, English only: small samples, so treat close numbers as ties.
Cozzo tested: reply time on real phone calls
Median time from the end of the caller’s speech to the agent’s first reply, on each vendor’s default model and voice (Oct 20, 2026). Places are out of 8 platforms tested.
- Retell AI: 1.04 s (tied 1 of 8, 11 calls)
- ElevenLabs Agents: 1.04 s (tied 1 of 8, 11 calls)
- Ultravox: 1.18 s (3 of 8, 11 calls)
- OpenAI Realtime API: 1.52 s (tied 5 of 8, 11 calls)
- LiveKit Agents: 1.52 s (tied 5 of 8, 11 calls)
- Vapi: 1.56 s (7 of 8, 11 calls)
Not tested yet: Gemini Live API.
See the full phone-call benchmark → Small samples, so treat close numbers as ties.
What to check before you choose
- Keep the agent brief. Agent speech is 1,200 tokens a minute at $64 per 1M, so on gpt-realtime-2.1 a minute of the agent talking is about 7.7¢ against about 1.9¢ for a minute of the caller talking. Short greetings and answers are the biggest saving you control.
- Protect the cache. Every reply re-sends the conversation so far, and cached input is $0.40 per 1M instead of $32. OpenAI's advice is to keep the session's history static: changing instructions or tools mid-call breaks the cache from that point on.
- Cap the history on long calls. OpenAI lets you set a token limit and a retention ratio for truncation (its example: keep 80% and limit input to 8,000 tokens after the instructions), which stops later turns from getting more and more expensive.
- Get it working, then try mini. gpt-realtime-2.1-mini is about a third of the audio price. OpenAI suggests testing on the full model first, then switching to mini once your prompt and tools are settled, because mini is weaker at following instructions and calling functions.
- Measure before you budget. OpenAI itself says Realtime costs are hard to estimate ahead of time. Run a few sample calls and read the token counts in each response.done event (or the Realtime Playground's Logs tab), then multiply by your call volume.
- Price the phone line too. OpenAI's SIP endpoint is free to connect, but you need a trunk from a carrier like Twilio, Telnyx or Bandwidth. On Twilio that's $1.15 a month for a US local number and $0.0085 a minute inbound.
- Compare whole minutes. Ultravox charges a flat $0.05 a minute including its model, and Vapi $0.05 a minute hosting plus your providers (about 8–13¢ a minute on Vapi's own calculator). On Cozzo's Oct 20 phone calls OpenAI's gpt-realtime-2.1 replied in a median 1.52 seconds over 11 calls, end to end including the phone line.
Questions people ask
How much does the OpenAI Realtime API cost per minute?
OpenAI doesn't publish a per-minute price for the Realtime API; it bills tokens. By OpenAI's token rules, a minute of caller audio is 600 tokens (about $0.019 on gpt-realtime-2.1) and a minute of agent speech 1,200 tokens (about $0.077). Because each reply re-sends the conversation, Cozzo's worked example of a 3-minute call comes to about 5¢ a minute if that history is cached and about 21¢ if it isn't. On gpt-realtime-2.1-mini it's about 1.6¢ to 6.5¢.
What does gpt-realtime-2.1 cost?
$32 per 1M audio input tokens ($0.40 cached) and $64 per 1M audio output tokens. Text is $4 in ($0.40 cached) and $24 out per 1M, and image input $5 per 1M. These are the rates on OpenAI's pricing page on October 25, 2026.
How much cheaper is gpt-realtime-2.1-mini?
Its audio is $10 per 1M tokens in ($0.30 cached) and $20 out, about a third of gpt-realtime-2.1's $32 and $64. Text is $0.60 in and $2.40 out per 1M. OpenAI says mini is weaker at instruction following and function calling.
How does OpenAI count audio tokens?
OpenAI's Realtime cost guide says audio in user messages is 1 token per 100 ms and audio in assistant messages 1 token per 50 ms, so 600 tokens per minute of caller audio and 1,200 per minute of agent audio. Special tokens add small variations.
Why do long Realtime calls cost more per minute?
Every reply sends the whole conversation so far back to the model as input, so later turns carry more input tokens than early ones. Caching makes the repeated part much cheaper ($0.40 per 1M instead of $32 on gpt-realtime-2.1), but OpenAI says caching is best-effort. Truncation settings and deleting or summarizing old turns keep it in check.
What is GPT-Live 1 and how is it priced?
GPT-Live 1 is OpenAI's voice session product priced by time: $0.05 a minute, billed per second, with no rounding up to a whole minute. Silence and backend work count as session time. The backend model and tools are billed separately at their own rates.
Does OpenAI charge for phone calls or SIP?
OpenAI says there's no charge for network bandwidth or connections. It doesn't sell phone numbers, though: you connect a SIP trunk from a carrier such as Twilio, Telnyx or Bandwidth, and that carrier bills the number and minutes. On Twilio a US local number is $1.15 a month and inbound calls $0.0085 a minute.
Is the OpenAI Realtime API cheaper than Vapi or Ultravox?
It can be, if most of the conversation hits the cache: Cozzo's worked example comes to about 5¢ a minute on gpt-realtime-2.1 and under 2¢ on mini, before text tokens and your carrier. Without cache hits it's about 21¢. Ultravox is a flat $0.05 a minute including its model, and Vapi is $0.05 a minute hosting plus providers, about 8 to 13¢ on Vapi's own calculator.
Has Cozzo measured what an OpenAI Realtime call costs?
Not as a measured bill. Our Round 4 cost figure for OpenAI was only an estimate, so on this page we quote OpenAI's published rates and our arithmetic from its token rules. We did measure speed: a median reply of 1.52 seconds over 11 phone calls on Oct 20, and 1.71 seconds in our Round 4 receptionist calls.
Ready to look closer? See OpenAI Realtime API’s full listing and price history → Or weigh the options in OpenAI Realtime API Alternatives Compared.
Not sure which fits? Answer a few quick questions in the finder →
Prices from vendors' official pricing pages, checked October 25, 2026. Not sponsored; order isn't a ranking.