Voice Agent Platforms

Vapi vs OpenAI Realtime API: priced, tested on real calls, and which to pick

Pricing and features, side by side. Last verified October 25, 2026. Spot a wrong price? support@cozzo.io

Cozzo’s tip: start with the billing row. It’s where the surprises hide.

Cozzo’s verdict. Short answer: they’re different kinds of product. Vapi is a voice agent platform: phone numbers, a dashboard, testing and your choice of speech-to-text, LLM and voice, for a $0.05-a-minute hosting fee plus the providers at cost. OpenAI’s Realtime API is a speech-to-speech model you build around yourself, billed per audio token, with your own phone carrier and server. On Cozzo’s Oct 20 phone calls they were about equally fast (1.56 s vs 1.52 s median reply). In our Round 4 receptionist test OpenAI passed 12 of 15 calls and Vapi 9, mostly because Vapi’s default agent said it couldn’t arrange callbacks. And you don’t have to pick just one: Vapi can run OpenAI’s realtime models too.

6 of 8 rows differ

Vapi and OpenAI Realtime API compared: type, billing style, starting price, billing, hipaa / baa, best for, key features, last verified
Detail	Vapi	OpenAI Realtime API
TypeDiffers	Voice Agent Platforms	Open-Source & Dev Frameworks
Billing styleSame	BothPay as you go
Starting priceDiffers	$0 to start; $0.05/min Vapi hosting, models billed pass-through at cost	gpt-realtime-2.1-mini: audio $10 in / $20 out per 1M tokens
BillingDiffers	Usage-based (per minute) + optional support package	Per token (audio, text, image); some per-minute products
HIPAA / BAADiffers	HIPAA add-on: $2,000/mo (DPA + BAA)	Not listed
Best forDiffers	Developers who want to choose their own LLM, transcriber, voice and telephony provider	Developers building speech-to-speech voice agents directly on OpenAI models
Key featuresDiffers	Bring-your-own model, transcriber and voice providers Unlimited agents on every plan Vapi telephony/SIP, WebSockets and WebRTC transport at no extra charge	Model works directly with audio, keeps conversation state and calls tools WebRTC in the browser or WebSocket on the server Barge-in, turn detection and tool use in one session
Last verifiedSame	BothOctober 25, 2026

Cozzo testedVoice agent latency on real phone calls

Vapi: 1.56 s median reply time (7 of 8, 11 calls)
OpenAI Realtime API: 1.52 s median reply time (tied 5 of 8, 11 calls)

Real phone calls, vendor defaults, Oct 20, 2026. Small samples, so treat close numbers as ties. See the method and limits.

Visit Vapi ↗Affiliate link

Cozzo may earn a commission if you sign up through some links. It never affects our rankings or test results. Disclosure

Measured on Cozzo’s test calls

Vapi: 1.56 s median reply time (11 phone calls); cost $0.042–0.051 per ~32 s call (about $0.08–0.09/min)
OpenAI Realtime API: 1.52 s median reply time (11 phone calls)

From our own phone calls on Oct 20, 2026, vendor defaults. Measured cost is what our calls cost, not a list price.

Cozzo-tested · Round 4 · October 22, 2026

AI receptionist test: Vapi vs OpenAI Realtime API

Calls passed out of 3 on each scenario, same dental-office prompt for both.

Vapi · 9/15 · 1.90 s‡ median reply

Book a new patient
Reschedule
Unknown insurance
Interruption
Dental emergency

OpenAI Realtime API · 12/15 · 1.71 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. 3 calls per scenario; small samples.

See the AI receptionist benchmark →All voice AI benchmarks →

Vapi Pricing Explained →

Vapi charges a $0.05-a-minute hosting fee and passes through the transcriber, language model and voice at the provider’s price with no markup. Its own calculator puts a Deepgram + OpenAI + ElevenLabs minute at about $0.08–$0.13, or $82–$129 for 1,000 minutes. Phone transport is free on Vapi numbers and SIP. OpenAI’s Realtime API has no platform fee or per-minute price: gpt-realtime-2.1 costs $32 per 1M audio tokens in ($0.40 cached) and $64 out, and gpt-realtime-2.1-mini $10 in and $20 out. OpenAI’s docs count caller audio at 1 token per 100 ms and agent audio at 1 token per 50 ms, and every reply re-sends the conversation so far as input, so the cost per minute depends on your calls. You bring a SIP carrier such as Twilio, and your server answers each call through a webhook.

Pick OpenAI’s Realtime API if you have developers and want the model itself, with no platform fee in between. It passed 12 of 15 Round 4 receptionist calls, took the callback message on every unknown-insurance call and told every emergency caller to get care now. It replied in a median 1.52 s on our Oct 20 phone calls (1.71 s in Round 4) and stopped within 0.10–0.14 s when a caller cut in (a reading under 0.2 s that we treat as unconfirmed). The catch: it read the callback number back wrong on all 3 booking calls, and you build call handling, tools, logging and testing yourself.

Pick Vapi if you want a working phone agent without building the plumbing: Vapi phone numbers, testing and simulation, 4 calls at once to start (10 on the $29 Core package, 30 on Pro, $10 a month for each extra line), a $2,000-a-month HIPAA add-on with a BAA, and any model you like, including OpenAI’s realtime models. Its default agent replied in a median 1.56 s on our Oct 20 phone calls and cost about $0.08–$0.09 a minute. Watch callbacks: in Round 4 it said it couldn’t arrange one on all 3 unknown-insurance calls, so tell your agent how to take messages.

Cozzo tip: try OpenAI’s voice inside Vapi first. Vapi can run OpenAI’s realtime models, so you get numbers and testing on day one, then decide if building it yourself saves enough to be worth it.

Sources: Vapi · OpenAI Realtime API

Vapi vs OpenAI Realtime API: prices and plans

	Vapi	OpenAI Realtime API
What you’re buying	A voice agent platform: transcriber, LLM and voice (or a speech-to-speech model), phone numbers, dashboard, tools, testing and simulation	A speech-to-speech model API (WebRTC, WebSocket or SIP). Your server handles call setup, tools, logging and testing
How you pay	Prepaid credits: $0.05/min Vapi hosting plus providers at their listed price, no markup. $5 free credits to start	Per token, charged as each reply is created: gpt-realtime-2.1 audio $32 in ($0.40 cached) / $64 out per 1M tokens, text $4 in / $24 out; gpt-realtime-2.1-mini audio $10 in ($0.30 cached) / $20 out
What one minute costs	About $0.082–$0.129 on Vapi’s calculator: hosting $0.05 + Deepgram $0.0095–$0.0099 + OpenAI LLM $0.0077–$0.0452 + ElevenLabs voice $0.0146–$0.0238	No per-minute price published. OpenAI’s token rules (Cozzo’s arithmetic): a minute of caller audio is 600 tokens ($0.0192 on gpt-realtime-2.1, $0.006 on mini) and a minute of agent speech 1,200 tokens ($0.0768, or $0.024 on mini), before the re-sent conversation and text tokens
Phone numbers and carrier	1 Vapi number included (5 on Core, 10 on Pro). Vapi telephony and SIP are free; your own carrier bills you (Vapi lists Twilio $0.008/min inbound, $0.014 outbound; Vonage $0.00814; Telnyx $0.0055)	No numbers sold. Point a SIP trunk (e.g. Twilio) at OpenAI’s SIP endpoint and your carrier bills you. OpenAI lists no charge for connections or bandwidth
Concurrent calls	4 to start, 10 on Core ($29/mo), 30 on Pro (10% of hosting, $999/mo minimum); extra lines $10 a month each	Not published as a number of calls
How calls are billed	Per minute an agent runs, from prepaid credits. Rounding rule not stated on the pricing page	Per reply: input (the whole conversation so far, cached where possible) plus output tokens. Silence isn’t billed when voice activity detection filters it out
Free testing	$5 free credits; testing and simulation included	No free credits published. OpenAI suggests measuring token use on sample sessions in its Realtime Playground
HIPAA / BAA	$2,000/mo add-on per organization, with a standard DPA and BAA	BAA on request, reviewed case by case (no enterprise agreement needed). /v1/realtime is HIPAA-eligible once your account has Modified Retention. No fee published
1,000 call minutes a month	$82–$129 on Vapi’s own calculator, before carrier fees and any package	Not published. Depends on how much each side talks and how long calls run
5,000 call minutes a month	About $409–$645 at the same per-minute range	Not published

Vapi

What you’re buying: A voice agent platform: transcriber, LLM and voice (or a speech-to-speech model), phone numbers, dashboard, tools, testing and simulation
How you pay: Prepaid credits: $0.05/min Vapi hosting plus providers at their listed price, no markup. $5 free credits to start
What one minute costs: About $0.082–$0.129 on Vapi’s calculator: hosting $0.05 + Deepgram $0.0095–$0.0099 + OpenAI LLM $0.0077–$0.0452 + ElevenLabs voice $0.0146–$0.0238
Phone numbers and carrier: 1 Vapi number included (5 on Core, 10 on Pro). Vapi telephony and SIP are free; your own carrier bills you (Vapi lists Twilio $0.008/min inbound, $0.014 outbound; Vonage $0.00814; Telnyx $0.0055)
Concurrent calls: 4 to start, 10 on Core ($29/mo), 30 on Pro (10% of hosting, $999/mo minimum); extra lines $10 a month each
How calls are billed: Per minute an agent runs, from prepaid credits. Rounding rule not stated on the pricing page
Free testing: $5 free credits; testing and simulation included
HIPAA / BAA: $2,000/mo add-on per organization, with a standard DPA and BAA
1,000 call minutes a month: $82–$129 on Vapi’s own calculator, before carrier fees and any package
5,000 call minutes a month: About $409–$645 at the same per-minute range

OpenAI Realtime API

What you’re buying: A speech-to-speech model API (WebRTC, WebSocket or SIP). Your server handles call setup, tools, logging and testing
How you pay: Per token, charged as each reply is created: gpt-realtime-2.1 audio $32 in ($0.40 cached) / $64 out per 1M tokens, text $4 in / $24 out; gpt-realtime-2.1-mini audio $10 in ($0.30 cached) / $20 out
What one minute costs: No per-minute price published. OpenAI’s token rules (Cozzo’s arithmetic): a minute of caller audio is 600 tokens ($0.0192 on gpt-realtime-2.1, $0.006 on mini) and a minute of agent speech 1,200 tokens ($0.0768, or $0.024 on mini), before the re-sent conversation and text tokens
Phone numbers and carrier: No numbers sold. Point a SIP trunk (e.g. Twilio) at OpenAI’s SIP endpoint and your carrier bills you. OpenAI lists no charge for connections or bandwidth
Concurrent calls: Not published as a number of calls
How calls are billed: Per reply: input (the whole conversation so far, cached where possible) plus output tokens. Silence isn’t billed when voice activity detection filters it out
Free testing: No free credits published. OpenAI suggests measuring token use on sample sessions in its Realtime Playground
HIPAA / BAA: BAA on request, reviewed case by case (no enterprise agreement needed). /v1/realtime is HIPAA-eligible once your account has Modified Retention. No fee published
1,000 call minutes a month: Not published. Depends on how much each side talks and how long calls run
5,000 call minutes a month: Not published

Prices read on vapi.ai/pricing, docs.vapi.ai, developers.openai.com/api/docs/pricing, OpenAI’s Realtime cost and SIP guides and OpenAI’s BAA and HIPAA help articles on Oct 25, 2026. Vapi rows use the stack in Vapi’s own calculator (Deepgram, an OpenAI LLM, ElevenLabs); the 5,000-minute row is that per-minute range × 5,000. OpenAI’s token figures are Cozzo’s arithmetic from OpenAI’s published audio-token rules, not a full call price. OpenAI also sells GPT-Live 1 voice sessions at $0.05 a minute, billed per second, with the backend model and tools charged on top; we haven’t tested GPT-Live.

What Cozzo measured on real calls

	Vapi	OpenAI Realtime API
Setup we tested	Vapi defaults: gpt-4o-mini, Deepgram nova-3, Vapi “Elliot” voice	gpt-realtime-2.1 over OpenAI’s SIP endpoint, default voice
Reply speed (Oct 20 phone calls)	1.56 s median (1.83 s p90), 11 calls	1.52 s median (2.00 s p90), 11 calls
Reply speed (browser link, no phone network)	1.15 s median	Not tested
Phone transcription errors (Oct 20)	2.2%	0.3% (our Whisper check: OpenAI gave us no transcript of the caller)
Round 4 receptionist test (dental script)	9 of 15 calls passed; 1.90 s median reply‡; callback number read back correctly 1 of 3 times; unknown insurance 0 of 3 (said it can’t arrange callbacks); emergency 2 of 3; interruption 3 of 3 (it had already stopped talking before the cut-in)	12 of 15 calls passed; 1.71 s median reply; callback number read back correctly 0 of 3 times (it said 415-555-0187, 412-550-187 and 412-555-0877); unknown insurance 3 of 3 but read the number back wrong at least once; emergency 3 of 3; stopped within 0.10–0.14 s when interrupted (under 0.2 s, treat as unconfirmed)
Round 5: gym	Passed every quote, booking, triage and after-hours call (said “901” before correcting to 911 on triage); stopped within 0.92 s; 1,780 ms median reply	Not tested
Round 5: career coaching	Passed every call; stopped within 0.84 s; 1,680 ms median reply	Not tested
Round 5: roofing & HVAC	Quote 2 of 2, booking 1 of 2, read-back 3 of 4, emergency passed, after-hours 0 of 1; 2,040 ms median reply	Not tested
Measured cost	$0.042–$0.051 per ~32 s call (about $0.08–$0.09/min), Vapi hosting plus pass-through model, transcription and voice	Not measured

Vapi

Setup we tested: Vapi defaults: gpt-4o-mini, Deepgram nova-3, Vapi “Elliot” voice
Reply speed (Oct 20 phone calls): 1.56 s median (1.83 s p90), 11 calls
Reply speed (browser link, no phone network): 1.15 s median
Phone transcription errors (Oct 20): 2.2%
Round 4 receptionist test (dental script): 9 of 15 calls passed; 1.90 s median reply‡; callback number read back correctly 1 of 3 times; unknown insurance 0 of 3 (said it can’t arrange callbacks); emergency 2 of 3; interruption 3 of 3 (it had already stopped talking before the cut-in)
Round 5: gym: Passed every quote, booking, triage and after-hours call (said “901” before correcting to 911 on triage); stopped within 0.92 s; 1,780 ms median reply
Round 5: career coaching: Passed every call; stopped within 0.84 s; 1,680 ms median reply
Round 5: roofing & HVAC: Quote 2 of 2, booking 1 of 2, read-back 3 of 4, emergency passed, after-hours 0 of 1; 2,040 ms median reply
Measured cost: $0.042–$0.051 per ~32 s call (about $0.08–$0.09/min), Vapi hosting plus pass-through model, transcription and voice

OpenAI Realtime API

Setup we tested: gpt-realtime-2.1 over OpenAI’s SIP endpoint, default voice
Reply speed (Oct 20 phone calls): 1.52 s median (2.00 s p90), 11 calls
Reply speed (browser link, no phone network): Not tested
Phone transcription errors (Oct 20): 0.3% (our Whisper check: OpenAI gave us no transcript of the caller)
Round 4 receptionist test (dental script): 12 of 15 calls passed; 1.71 s median reply; callback number read back correctly 0 of 3 times (it said 415-555-0187, 412-550-187 and 412-555-0877); unknown insurance 3 of 3 but read the number back wrong at least once; emergency 3 of 3; stopped within 0.10–0.14 s when interrupted (under 0.2 s, treat as unconfirmed)
Round 5: gym: Not tested
Round 5: career coaching: Not tested
Round 5: roofing & HVAC: Not tested
Measured cost: Not measured

Small samples (1–3 calls per scenario per vendor in Rounds 4 and 5, 11 calls each on Oct 20), so treat close numbers as ties. OpenAI’s Oct 20 calls were part of Round 3 the same afternoon, and OpenAI wasn’t in Round 5. ‡ Vapi’s 1.90 s is end-to-end on an untuned default setup, not comparable to Vapi’s published figures, which leave out endpointing and network time. We tested Vapi on its default pipeline, not on OpenAI’s realtime models inside Vapi. Full method and limits are on each benchmark page.

Questions

Is Vapi cheaper than the OpenAI Realtime API?

It depends on your calls, because OpenAI publishes no per-minute price. Vapi is $0.05 a minute plus providers at cost, about $0.08–$0.13 a minute on its own calculator, and our default Vapi calls cost about $0.08–$0.09 a minute. OpenAI bills tokens: on gpt-realtime-2.1 a minute of caller audio is about $0.019 and a minute of agent speech about $0.077 by OpenAI’s token rules, and each reply re-sends the conversation so far as input, so long calls cost more per minute. The mini model is about a third of those audio rates. We haven’t measured an OpenAI bill per call.

Which is faster, Vapi or OpenAI Realtime?

About the same over the phone. On Cozzo’s Oct 20 phone calls the median reply was 1.56 s for Vapi’s default setup and 1.52 s for OpenAI’s gpt-realtime-2.1 over SIP. In our Round 4 receptionist calls OpenAI was quicker (1.71 s vs 1.90 s), though Vapi’s figure there is on an untuned default setup.

Which handled receptionist calls better in Cozzo’s tests?

OpenAI, in Round 4: 12 of 15 calls passed against Vapi’s 9. Vapi’s default agent said it couldn’t arrange callbacks on all 3 unknown-insurance calls and missed 1 of 3 emergency calls, while OpenAI took every callback message and handled every emergency. Neither got the callback number right on booking calls (Vapi 1 of 3, OpenAI 0 of 3). OpenAI wasn’t in Round 5; Vapi passed every gym and coaching call there.

Can Vapi use OpenAI’s Realtime models?

Yes. Vapi’s docs list GPT Realtime 2 (recommended for new assistants), GPT Realtime Mini and an older realtime model in its model picker, with OpenAI voices such as marin and cedar. You pay Vapi’s $0.05-a-minute hosting plus OpenAI’s cost with no markup. We haven’t tested that setup.

Do Vapi and OpenAI support HIPAA?

Both can. Vapi sells HIPAA-eligible data handling as a $2,000-a-month add-on per organization, with a standard DPA and BAA. OpenAI signs BAAs for its API on request, reviewed case by case, and lists /v1/realtime as HIPAA-eligible once your account has Modified Retention. OpenAI doesn’t publish a fee for it.

Can I use my own phone number?

Yes on both. Vapi includes 1 phone number (5 on Core), its own telephony and SIP are free, and it works with your Twilio, Vonage or Telnyx number, which that carrier bills. OpenAI doesn’t sell numbers: you point a SIP trunk from a carrier such as Twilio at OpenAI’s SIP endpoint, and your server accepts each call through a webhook.

How many calls can each handle at once?

Vapi allows 4 at once on usage alone, 10 on its $29-a-month Core package and 30 on Pro, and sells extra lines at $10 a month each. OpenAI doesn’t publish a concurrent-call number for the Realtime API.

Keep comparing

Related comparisons

Vapi vs Retell AI→Vapi vs Bland AI→Synthflow vs Vapi→ElevenLabs Agents vs Vapi→