OpenAI Realtime API alternatives in 2027 for voice agents
Seven OpenAI Realtime API alternatives for teams whose audio token bill, missing phone line or single model setup has become the bottleneck. PyAI Omni and Retell AI ship with phone numbers, Gemini Live is the nearest native audio model, and LiveKit hands you the source code.

On this page
Why teams look for OpenAI Realtime API alternatives
gpt-realtime handles transcription, reasoning and voice inside one WebRTC or WebSocket session, and few teams leave over sound quality. The four reasons below come from 2026 developer write ups by AssemblyAI and Dasha.
Audio tokens add up on long calls
Each turn resends the conversation so far, so AssemblyAI puts real bills at 2 to 5 times the $32 and $64 headline on longer calls, and audio tokens already cost several times more than text.
No phone numbers or SIP
Realtime stops at the media stream, with no PSTN, SIP trunk or call routing, so a support line needs Twilio or LiveKit SIP added and billed separately.
One model, no swaps
Dasha calls it a black box: recognizer, LLM and voice are all gpt-realtime, so you cannot replace just the voice or tune turn taking on its own.
Barge in and noise
AssemblyAI notes that background noise can trigger barge in, so a cough stops the agent mid sentence, and noisy audio sometimes yields words the caller never said.
OpenAI Realtime API alternatives at a glance
Prices come from vendor pricing pages read between 20 and 30 October 2026. Where a vendor publishes only parts of the bill, the third party estimate is named.
| Tool | Built for | Features | Free option | Price (2026) | Telephony |
|---|---|---|---|---|---|
| PyAI | Teams wanting managed agents | Omni speech to speech with Hear, Speak and Cast on one OpenAI-compatible API, telephony included | $50 credit, no credit cardTest keys are not charged | $0.05/min + $0.01 telephony | Included |
| Gemini Live API | Google native builders | Native audio in and out from Google, token based with no built in calling | Free tier in AI Studio | Token based, audio output the main cost | None, add a carrier |
| Vapi | Developers wiring providers | A managed layer where you bring your own models, voices and carriers | $5 credit | $0.05/min hosting, providers at cost | Twilio, Telnyx or SIP |
| Retell AI | Self serve voice teams | A self serve builder with base call time plus a separate model charge | $10 credit | $0.07/min base, LLM extra | Numbers at $0.015/min |
| Deepgram Voice Agent | Voice agent API devs | Transcription and voices joined in one agent endpoint, carrier separate | $200 credit | $0.075/min Standard agent | Separate carrier |
| ElevenLabs Agents | Lifelike voice shops | An agent layer over the ElevenLabs voice library with instant cloning | Free tier credits | $0.08/min call time, LLM extra | Via integration |
| LiveKit | Open framework teams | An open source real time framework with native SIP calling | Build plan, 1,000 minutes | Ship from $50/month, then $0.01/min | Native SIP |
| OpenAI Realtime (for comparison) | Direct model integrators | A speech to speech model over the Realtime API, with no telephony | Pay as you goNo telephony | $32 in, $64 out per 1M audio tokens | Not offered |
OpenAI Realtime API cost vs alternatives
Because gpt-realtime meters audio tokens and resends context every turn, a long call costs more per minute than a short one. The chart puts 10,000 call minutes a month on each platform that publishes a per minute price. For other volumes, the PyAI voice AI cost calculator takes your own minutes and call length.
PyAI vs OpenAI Realtime, 10,000 minutes a month
10,000 Omni minutes cost $600 a month, number included
Each bar is the monthly bill for 10,000 call minutes, built from vendor rates and vendor estimators. The right column sets each one against the Realtime flagship range.
- PyAI OmniTelephony includedSave up to $500
$600 a month: 10,000 minutes at $0.05 plus $0.01 for the number, with speech and reasoning in the same rate
- Deepgram Standard agentDeveloper APIInside the range
$750 a month on the $0.075 Standard agent, which bundles the LLM but leaves the phone carrier to you
- ElevenLabs AgentsNatural speechBefore LLM costs
$800 a month in call time alone at $0.08, before the LLM bill and a carrier
- OpenAI RealtimeThe tool being replacedBaseline
$600 to $1,100 a month if third party estimates of $0.06 to $0.11 hold, rising with call length and before a carrier
- VapiYour own providersCan cost more
$820 to $1,290 a month from the $82 to $129 per 1,000 minutes Vapi estimates, with the carrier billed by your provider
- Gemini Live and LiveKitModel or frameworkNot comparable
Priced by tokens or self hosted. Gemini Live bills audio tokens with no telephony, and LiveKit is an open source framework you run
Rates read 20 to 30 October 2026. The Realtime range is a third party conversion of audio token prices to minutes and excludes any carrier.
OpenAI Realtime API competitors reviewed
Seven tools, ordered by how much of the stack each one takes off your hands. PyAI, Retell AI and ElevenLabs run the agent for you, Vapi and Deepgram leave vendor choices with developers, Gemini Live mirrors the Realtime model, and LiveKit hands over the code.
PyAI
Built for phone agents that need a flat per minute bill and a line on the same account
PyAI is voice infrastructure. The Omni engine, like gpt-realtime, hears, reasons and speaks inside one session, but Omni charges by the minute and comes with managed phone numbers. PyAI replaces the OpenAI Realtime API for phone agents without a token meter or a second carrier contract. PyAI is not a raw model you orchestrate yourself and not an open framework, so teams that want to swap each component should look at LiveKit or Pipecat.
Omni costs $0.05 a minute plus $0.01 for the phone line and metered per minute. Tests run by PyAI measure about 196 ms to detect the end of a turn, a 390 ms median to the first reply audio, and a 4.76% word error rate on 8 kHz phone audio. The API is OpenAI compatible, so a Realtime client mostly needs a new base URL and endpoint changes.
Since PyAI is a young product, its G2 listing has no reviews yet.
PyAI features
- Omni
- Speech to speech engine that holds its reply through half finished thoughts and digit strings, then answers on the same phone line.
- Per minute billing
- $0.05 a minute plus $0.01 for the number, so a 12 minute call costs exactly 12 times a 1 minute call.
- OpenAI compatible
- Point an existing OpenAI client at the PyAI base URL instead of rewriting the integration.
- Turn detection
- Reads words and meaning together instead of a silence timer, aimed at the false barge in AssemblyAI flags on Realtime.
- Trace and Recap
- Every call gets a quality and compliance score plus a written summary, so a migration is judged on real calls.
PyAI pricing
Omni
$0.05/min
Plus $0.01 phone line
Agents
$0.08/min
No code builder
Credit
$50
No credit card
Each call is metered by the minute, test keys are never charged, and every new account opens with $50 of free credit and no credit card. Sample voices play on PyAI Speak.
PyAI pros and cons
Pros
- One rate of $0.05 a minute plus $0.01 for the number covers speech, reasoning and the phone number, so a 10,000 minute month is a known $600 instead of a token estimate.
- Managed numbers come on the same account, removing the Twilio or SIP layer every Realtime phone agent needs.
- OpenAI compatible requests let a team leaving gpt-realtime keep most of its client code and prompt work.
- A 390 ms median reply and a 94.3 out of 100 agent quality score in PyAI tests keep Omni inside normal conversational pace on calls.
Cons
- No raw model or open source code to self host, so teams that must run every layer on their own servers should pick LiveKit or Pipecat.
- Omni bundles recognizer, LLM and voice, so plugging in one specific vendor for a single layer is not possible the way Vapi allows.
- No G2 reviews exist yet, and the 196 ms and 390 ms timings are PyAI measurements in a US region, not an outside benchmark.
PyAI reviews
G2 shows no PyAI reviews to date, so this tab is left blank rather than filled with made up quotes. The free sandbox key carries $50 of credit, enough to rerun a Realtime call flow on Omni and compare transcripts.
Choose PyAI if your Realtime agents live on the phone and you want one per minute bill with the number included. At 10,000 minutes a month that is $7,200 a year.
Google Gemini Live
Built for a Google native audio model that listens and talks in one stream
Google Gemini Live is the closest swap in shape: a native audio model that listens and talks in a single streaming session, so a Realtime prototype maps onto it almost one to one.
Google also meters by token, with audio output making up most of the bill, and AI Studio offers a free tier for testing.
The tradeoff mirrors Realtime rather than removing it: Gemini Live carries no phone numbers, SIP or call routing, so a team still bolts on Twilio or another carrier to answer a call, and the token meter still makes a long call cost more per minute than a short one. Gemini Live fits teams already inside Google Cloud who want to trial a second native audio model against gpt-realtime, keep telephony on infrastructure they already run, and are weighing model quality rather than hunting for an all in per minute phone agent.
Google Gemini Live features
- Native audio
- Gemini Live listens and replies inside one streaming session using a native audio model.
- Near one to one port
- Gemini Live keeps the single session shape, so a Realtime prototype moves mostly by changing the SDK.
- AI Studio tier
- Gemini Live can be tried free in Google AI Studio before any paid usage.
- Google billing
- Gemini Live bills through the Gemini and Google Cloud stack, useful for teams already paying Google.
Against OpenAI Realtime API's pain points: Google Gemini Live
Gemini Live is the like for like swap, which also carries over most Realtime pain points.
- Call cost: Gemini Live also meters by token with audio output as the largest share, so long calls stay expensive in the same way
- Telephony: Gemini Live has no PSTN or SIP layer, so Twilio, LiveKit or another carrier is still billed on top
- Model swaps: Gemini Live runs one native audio model end to end, so the voice cannot be replaced on its own any more than in gpt-realtime
- Barge in: Gemini Live barge in behavior on noisy lines is not covered in the sources checked, so replay real call audio in AI Studio first
- Event handling: Gemini Live keeps a streaming session model, so the developer still owns session and response event code
- Free testing: Gemini Live offers the AI Studio free tier, enough to compare voice and latency with Realtime at no cost
Google Gemini Live pros and cons
Pros
- Same single session design as gpt-realtime, so moving a prototype changes the vendor SDK more than the agent logic.
- A free tier in Google AI Studio lets a team compare voice and latency against Realtime before paying anything.
- Google Gemini Live sits inside the wider Gemini and Google Cloud stack, handy for teams already billing through Google.
Cons
- Billing is still per token with audio output the largest cost, so the long call problem that pushes teams off Realtime stays.
- No PSTN or SIP of its own, so a phone agent needs Twilio, LiveKit or another carrier added and paid separately.
Choose Google Gemini Live if you like the single model design of Realtime but want to test Google models, and you already handle telephony elsewhere.
Vapi
Built for orchestration where you choose the transcriber, LLM and voice yourself
Vapi orchestrates a voice agent from parts you choose: a transcriber such as Deepgram, an LLM such as GPT, and a voice such as ElevenLabs. Vapi's layered design buys the per layer control a single gpt-realtime session cannot give.
Vapi bills $0.05 a minute for hosting and passes provider costs through at cost. Vapi's own estimator puts 1,000 minutes at $82 to $129. Signup includes $5 of credit, and extra concurrent lines cost $10 a month each.
The cost of that control is a bill assembled from several vendors, since speech and model are passed through at provider rates on top of the hosting fee, so the all in minute lands above a single bundled rate and moves with the components chosen. Vapi fits developers who want to mix best of breed speech, model and voice and tune each layer, and who accept managing several provider accounts and a variable per minute total in return for that flexibility.
Vapi features
- Pick each layer
- Vapi assistants combine a chosen recognizer, model and voice, for example Deepgram, GPT and ElevenLabs.
- Hosting rate
- Vapi takes $0.05 per minute for hosting, and each provider is charged on at its own price.
- Carrier choice
- Vapi routes calls over Twilio, Telnyx or a customer SIP trunk without a transport fee of its own.
- Canvas and API
- Vapi offers a visual canvas for operations staff beside an API for developers.
- Extra lines
- Vapi charges $10 a month for each concurrent line past the included ones.
- Plan floors
- Vapi Pro carries a $999 monthly minimum, and HIPAA adds $2,000 a month.
Against OpenAI Realtime API's pain points: Vapi
Vapi breaks the single Realtime session into parts a team can replace one by one.
- Call cost: Vapi bills per minute rather than per token resent each turn, and its own estimate places 1,000 minutes between $82 and $129
- Telephony: Vapi connects to Twilio, Telnyx or your own SIP trunk, closing the phone gap Realtime leaves open
- Model swaps: Vapi lets a team keep GPT for reasoning and move only speech recognition to a cheaper engine
- Barge in: Vapi lets the team pick the transcriber, so a recognizer that copes better with noise can be tested without changing the LLM
- Event handling: Vapi wraps call logic in a canvas and API, so the 30 plus raw Realtime event types are not hand coded
- Free testing: Vapi includes $5 of signup credit, a small allowance once provider costs are added
Vapi pros and cons
Pros
- Each layer can be swapped on its own, so a team can keep GPT for reasoning while moving speech to a cheaper recognizer.
- Calls run over Twilio, Telnyx or your own SIP trunk, and Vapi does not charge for its own transport.
- A visual canvas plus an API suits both developers and the operations staff who edit call flows.
Cons
- Bills arrive from Vapi plus every provider, and the $82 to $129 per 1,000 minute estimate sits well above one flat rate.
- HIPAA adds $2,000 a month and the Pro plan carries a $999 monthly minimum, steep for a small team.
- G2 lists only 3 Vapi reviews at 4.2 out of 5, too few to judge support quality.
Vapi reviews
Vapi holds 4.2 on G2 from 3 reviews. Vapi on G2 ↗
A reviewer highlights: "Vapi AI is very easy to use and the integration is straightforward. The initial setup was very easy, which made the whole process smooth." (Bappy R., Small-Business, G2 May 2026)
On pricing: "Some limits as to the functionality, but mostly it is a superior and affordable product." (Steve M., Small-Business, G2 Mar 2025)
Choose Vapi if you want to pick each speech and LLM vendor while Vapi handles orchestration and scaling.
Retell AI
Built for teams wanting a no contract phone agent with call analytics and its own numbers
Retell AI is a finished agent product rather than a raw API: a drag and drop flow builder, warm transfers, call analytics and its own phone numbers come in the box. On build effort it sits at the opposite end from Realtime.
Retell lists a $0.07 a minute base, $0.055 for voice infrastructure plus $0.015 for standard speech, with the LLM on top and numbers at $0.015 a minute. Premium ElevenLabs voices lift the speech share to $0.040. New accounts get $10 of credit, and G2 rates it 4.8 out of 5 across 2,639 reviews.
The flip side of that finished product is less low level control than a raw API like Realtime gives, because the flow builder and managed stack set the guardrails and the speech, model and voice choices sit inside Retell's own tiers. Retell AI fits teams that want a working phone agent with analytics and its own numbers in days rather than building and maintaining code against an audio stream.
Retell AI features
- Flow builder
- Retell ships a drag and drop flow builder for phone agents.
- Warm transfer
- Retell hands callers to a human with warm transfer as part of the standard product.
- Phone numbers
- Retell numbers cost $0.015 a minute, or a team can bring its own SIP trunk.
- Base rate
- Retell lists $0.07 a minute, made of $0.055 voice infrastructure and $0.015 standard speech, before the LLM.
- Premium voices
- Retell raises the speech share to $0.040 a minute with premium ElevenLabs voices.
Against OpenAI Realtime API's pain points: Retell AI
Retell is the finished product at the opposite end from a raw Realtime API.
- Call cost: Retell charges by the minute, and vendor and third-party estimates from 24 June 2026 put a fully loaded Retell call between $0.13 and $0.31 once the LLM, voice and line are added
- Telephony: Retell sells its own numbers at $0.015 a minute and also accepts an existing SIP trunk
- Model swaps: Retell lets the team choose the LLM and the voice tier separately, from standard speech to premium ElevenLabs
- Barge in: Retell does not publish a noise or false barge in figure, so test a noisy line during the $10 credit period
- Event handling: Retell keeps session events behind its builder, trading fine grained event control for less code to maintain
- Free testing: Retell gives $10 of signup credit with no platform fee, so a pilot carries no monthly commitment
Retell AI pros and cons
Pros
- With 2,639 G2 reviews at an average of 4.8, Retell has more user feedback than any other tool in this list.
- A flow builder, warm transfer to a human and post call analytics are part of the standard product, not paid extras.
- Numbers cost $0.015 a minute, or teams bring their own SIP trunk to keep an existing carrier.
- No platform fee and $10 of signup credit, so a pilot carries no monthly commitment.
Cons
- Costs stack per component: the $0.07 base rises once an LLM, a premium voice and the line are added, and a fully loaded Retell call lands, per vendor and third-party estimates dated 24 June 2026, at $0.13 to $0.31 a minute.
- G2 reviewers mention missing features and a learning curve, and a builder gives less low level control than streaming events from a model API.
Retell AI reviews
Retell AI holds 4.8 on G2 from 2,639 reviews. Retell AI on G2 ↗
A reviewer highlights: "Retell AI is easy to use, highly customizable, and makes it simple to build realistic AI voice agents for lead qualification, customer support and appointment booking." (Verified user, Small-Business, G2 2026)
Another flags: "The main downside is the learning curve for advanced workflows and the time required to thoroughly test conversation flows so they feel consistently natural." (Verified user, Small-Business, G2 2026)
On pricing: "The pricing can also become expensive when scaling or doing many test calls during development." (Verified user in computer software, G2 Jun 2026)
Choose Retell AI if you would rather configure a finished phone agent with analytics than write code against an audio stream.
Deepgram Agent API
Built for developers who want Nova transcription and Aura speech in one agent rate
Deepgram Voice Agent puts Nova transcription, an LLM and Aura speech behind one agent endpoint, which makes it the developer option that feels closest to calling Realtime.
Standard costs $0.075 a minute on pay as you go and $0.068 on Growth. Supplying both the LLM and the voice drops the rate to $0.050 pay as you go and $0.041 on Growth. Signup credit is $200 and never expires. Phone calls still need a carrier you add.
The gap against a packaged phone agent is that telephony stays separate: Deepgram serves recognition, model and voice through one endpoint but ships no phone numbers, so a carrier such as Twilio still sits in front for live calls and the per minute total depends on whether you bring your own model and voice. Deepgram Voice Agent fits developers who already run Nova transcription and want the model and Aura voice folded into one agent rate, and who are comfortable wiring their own phone layer.
Deepgram Agent API features
- One endpoint
- Deepgram Voice Agent serves recognition, reasoning and Aura voices through one agent connection.
- Standard rate
- Deepgram prices Voice Agent Standard at $0.075 per minute without commitment and $0.068 per minute on Growth.
- Bring your own
- Deepgram drops to $0.050 pay as you go and $0.041 on Growth when the buyer supplies both the LLM and the voice.
- Lasting credit
- Deepgram signup credit of $200 does not expire and needs no credit card.
- Growth concurrency
- Deepgram Growth caps simultaneous agent calls at 60 for North America, and 45 for Europe, Australia and India.
- Code first
- Deepgram Voice Agent is built for developers and has no visual flow builder.
Against OpenAI Realtime API's pain points: Deepgram Agent API
Deepgram feels closest to calling Realtime, with a flat minute rate instead of growing tokens.
- Call cost: Deepgram charges a flat $0.075 a minute including the LLM and voice, so a long call costs minutes, not a resent transcript
- Telephony: Deepgram has no numbers of its own, so 10,000 Standard minutes at $750 still leave the carrier bill out
- Model swaps: Deepgram lets the buyer bring a separate LLM, a separate voice or both, with the rate falling to $0.050 pay as you go and $0.041 on Growth
- Barge in: Deepgram puts Nova transcription at the front of the agent, though noise triggered barge in is not measured in the sources checked
- Event handling: Deepgram still expects code for every flow, since operations staff get no visual editor
- Free testing: Deepgram $200 of credit covers more than 2,600 Standard minutes, the largest allowance on this page
Deepgram Agent API pros and cons
Pros
- The $0.075 Standard rate already includes the LLM and voice, simpler to forecast than audio tokens that grow each turn.
- $200 of signup credit with no expiry is the largest free allowance here, enough for more than 2,600 Standard minutes.
Cons
- No phone numbers of its own, so the $750 a month that 10,000 Standard minutes cost leaves out the carrier.
- Growth caps Voice Agent concurrency at 60 calls in North America and 45 in Europe, Australia and India, a ceiling busy support lines can reach.
- Built for developers writing code, with no visual builder for operations staff to edit flows.
Deepgram Agent API reviews
Deepgram on G2 (company rating): 4.6 from 482 reviews. The quotes below are about Deepgram's core speech product, not the Agent API. Deepgram on G2 ↗
A reviewer highlights: "Deepgram is really a developer-friendly solution and is extremely fast if you compare it against other vendors like OpenAI Whisper and Speechmatics." (Diego P., Enterprise, G2 Aug 2026)
Another flags: "Although the accuracy is very good, it can sometimes be a bit difficult to adjust the models for very specific cases or when there is technical jargon." (Verified user, Small-Business, G2 Aug 2026)
On pricing: "It supports word recognition and speaker diarization, and it also has a very good price offering comparatively." (Akshay M., Mid-Market, G2 Aug 2026)
Choose Deepgram Agent API if you already trust Deepgram transcription and want the LLM and voice priced into one agent rate.
ElevenLabs Agents
Built for agents where voice realism outranks the per minute price
ElevenLabs Agents puts the ElevenLabs voice library, the main reason teams choose the brand, inside a managed conversational agent, so the selling point is how human the caller hears it rather than raw model access.
Call time costs $0.08 a minute on every plan, down from $0.10, while the LLM is billed separately, from about $0.0005 a minute on GPT-5 Nano upward. Subscriptions span $6 on Starter to $990 on Business. G2 rates ElevenLabs 4.5 out of 5 across 1,260 reviews.
The tradeoff is price against realism: at $0.08 a minute plus a separate LLM charge, ElevenLabs Agents sits at the higher end of the packaged options, so the voice quality has to earn that premium against a cheaper bundled rate. ElevenLabs Agents fits brands where a caller hearing a natural voice outranks the per minute total, and teams that already build on the ElevenLabs library and want a managed agent around it.
ElevenLabs Agents features
- Voice library
- ElevenLabs Agents runs the ElevenLabs voice library inside a managed conversational agent.
- Call time
- ElevenLabs Agents charges $0.08 a minute of call time on every plan, down from $0.10.
- LLM billing
- ElevenLabs charges model usage on its own line, starting near $0.0005 per minute with GPT-5 Nano.
- Silence discount
- ElevenLabs bills silence over 10 seconds at about 95% off.
Against OpenAI Realtime API's pain points: ElevenLabs Agents
ElevenLabs Agents wins on how the voice sounds, with call time and the LLM billed apart.
- Call cost: ElevenLabs bills $0.08 a minute plus the LLM, and vendor and third-party estimates from 24 June 2026 put a fully loaded call near $0.10 to $0.15 a minute
- Telephony: ElevenLabs Agents has no carrier in its call rate, so a phone line is a separate bill on top of $800 per 10,000 minutes
- Model swaps: ElevenLabs separates the LLM from the voice, so a cheap model such as GPT-5 Nano can sit behind a premium voice
- Barge in: ElevenLabs noise and interruption handling is not described in the sources checked, so test with real background sound
- Event handling: ElevenLabs Agents is a managed agent, so the team configures behavior rather than handling raw session events
- Free testing: ElevenLabs paid plans start at $6 on Starter, and no free agent allowance is stated on this page
ElevenLabs Agents pros and cons
Pros
- A 4.5 G2 rating across 1,260 reviews, with quality and speed the most praised points, gives ElevenLabs the strongest voice reputation here.
- Silence longer than 10 seconds is billed at about 95% off, which helps on calls with long holds or lookup pauses.
Cons
- The $0.08 call time comes before the LLM and a carrier, so 10,000 minutes start at $800, and a fully loaded call lands, per vendor and third-party estimates dated 24 June 2026, at $0.10 to $0.15 a minute.
- Calls past plan concurrency bill as burst minutes at $0.16, double the normal rate.
ElevenLabs Agents reviews
ElevenLabs on G2 (company rating): 4.5 from 1,260 reviews. The ElevenLabs review quotes cover the core ElevenLabs product rather than Agents specifically. ElevenLabs on G2 ↗
A reviewer highlights: "The absolute best part is just how genuinely human the voices sound: the natural pauses, the breathing, and the way emphasis shifts." (Pradumya K., Enterprise, G2 Oct 2026)
Another flags: "The main pain point is how it handles dense legal terminology, citations and Latin maxims. The model frequently stumbles over citations." (Shreyansh P., Small-Business, G2 Oct 2026)
On pricing: "The best part about ElevenLabs is the free credit thing they have." (Rashi R., Small-Business, G2 Oct 2026)
Choose ElevenLabs Agents if a caller hearing a natural voice matters to your brand more than the per minute total.
LiveKit
Built for an Apache 2.0 agent framework with SIP built in and an optional cloud
LiveKit Agents is an Apache 2.0 framework with Python and Node SDKs, native SIP and an optional managed cloud. Teams keep the code and choose every model without running all the media servers themselves.
Build is free with 1,000 agent minutes and one US number, Ship costs $50 a month for 5,000 minutes then $0.01 each, and Scale starts at $500. LiveKit pricing shows a sample phone agent at $0.0479 a minute across session, telephony, LLM, STT, TTS and observability.
Ownership is the trade: the Apache 2.0 code and free model choice come with running or paying for the media and SIP infrastructure and wiring each provider yourself, work a managed engine hides behind one rate. LiveKit fits engineering teams that want to keep the code and swap any model today, and that are ready for the managed cloud once traffic outgrows a self hosted setup.
LiveKit features
- Apache 2.0 code
- LiveKit Agents is released under Apache 2.0 with Python and Node SDKs.
- Native SIP
- LiveKit includes SIP, with the first US number free and extra numbers at $1 a month.
- Plan ladder
- LiveKit Build is free with 1,000 minutes, Ship is $50 for 5,000 minutes and Scale starts at $500.
- Overage
- LiveKit charges $0.01 a minute after the bundled minutes are used.
- Sample agent
- LiveKit's pricing page totals an example phone agent near $0.0479 per minute, adding up session, line, model, speech and observability costs.
Against OpenAI Realtime API's pain points: LiveKit
LiveKit gives full control of every layer Realtime locks together, with more parts to manage.
- Call cost: LiveKit bills minutes and chosen providers, and the sample phone agent near $0.0479 a minute shows a low all in figure
- Telephony: LiveKit includes SIP and one free US number, a layer the Realtime API lacks entirely
- Model swaps: LiveKit leaves STT, LLM and TTS choice to the team, though each extra vendor adds a bill and a failure point
- Barge in: LiveKit Cloud includes turn detection free, while self hosted agents pay $0.50 per 1,000 requests past the allowance
- Event handling: LiveKit SDKs give the code to the team, so event upkeep moves into a framework the team can read and change
- Free testing: LiveKit Build includes 1,000 agent minutes and one US number at no charge
LiveKit pros and cons
Pros
- Source code under Apache 2.0, so the agent can move to your own servers whenever compliance or cost demands it.
- Native SIP with US numbers at $1 a month after the first free one, something Realtime does not offer at all.
- The LiveKit sample phone agent at $0.0479 a minute is among the lowest all in figures here, with low cost models chosen.
Cons
- You choose and pay each STT, LLM and TTS vendor, so bills and failure points multiply with every provider added.
- TRTC, a competitor, points out that past Build you pay $50 or $500 a month whether or not the bundled minutes get used.
- Turn detection is free on LiveKit Cloud, but self hosted agents on Ship or Scale pay $0.50 per 1,000 requests once the development allowance runs out.
- G2 shows just 2 LiveKit reviews, at 4.8, so peer feedback is thin.
LiveKit reviews
LiveKit holds 4.8 on G2 from 2 reviews. LiveKit on G2 ↗
A reviewer highlights: "I really appreciate LiveKit's ease of integration and its scalability, which is extremely helpful for our use case. We rely heavily on its voice agents stack, along with the noise suppression and echo cancellation modules. It's great that LiveKit supports STT, TTS, LLM, custom LLMs, and SIP telephony integration." (Viren S., User, G2 September 2026)
Choose LiveKit if your engineering team wants open source code and SIP today, with a hosted cloud ready for when traffic grows.
OpenAI Realtime API alternatives by job and use case
Use the tabs to match a Realtime replacement to the job, the build style and the billing model you can live with.
Which tool fits which job
| Job | Pick | Why |
|---|---|---|
| Managed phone agents | PyAI, Retell AI | PyAI line in the rate. Retell numbers $0.015/min |
| Single model like Realtime | Gemini Live | Closest like for like |
| Provider control | Vapi, Deepgram | Choose each vendor |
| Voice realism first | ElevenLabs Agents | 4.5 on G2, praised for quality |
| Open source | LiveKit, Pipecat | Self host, full control |
| Flat rate with a number | PyAI Omni | $0.05 a minute plus $0.01 for the line |
Raw model vs managed agent
| Raw model (Realtime, Gemini Live) | Managed (PyAI, Retell) | |
|---|---|---|
| Scope | Speech and reasoning | Agent plus telephony |
| Telephony | Add a carrier | PyAI line included. Retell $0.015/min |
| Pricing | Per token | PyAI flat per minute. Retell stacks per part |
| Control | One model, few knobs | Configured, swappable |
| Best for | Custom builds | A phone agent fast |
Pricing models compared
| Model | Who uses it | What to watch |
|---|---|---|
| Per token | OpenAI Realtime, Gemini Live | Cost climbs with call length |
| Flat per minute | PyAI | $0.01 telephony on top of $0.05 |
| Per minute by component | Retell | $0.13 to $0.31 fully loaded |
| Engine plus components | Vapi, Deepgram | Costs stack by part |
| Open source | LiveKit, Pipecat | You run and scale it |
Free and open source OpenAI Realtime alternatives
Two open source frameworks cover most self hosted Realtime replacements: LiveKit Agents under Apache 2.0 and Pipecat under BSD-2, with Pipecat Cloud at $0.01 per agent minute for a hosted version. An open speech-to-speech model and an open multimodal model go further as open speech to speech models. Each trades the token meter for servers, provider bills and engineering hours. Managed trials cost nothing up front: PyAI gives $50 with no credit card, Deepgram $200, Retell AI $10 and Vapi $5.
Before moving anything, replay your highest volume call flow on two finalists, because barge in and response time behave differently with real callers on a phone line than in a browser demo.
When the OpenAI Realtime API is still the right pick
Staying on gpt-realtime makes sense for in app or browser voice where no phone line is needed, where short sessions keep token bills small, and where the team already builds on OpenAI. gpt-realtime-mini, at $10 and $20 per million audio tokens, also trims cost for simple flows. Switching pays off most for phone agents with long calls, high monthly volume or a need to change one layer.
How to migrate off the OpenAI Realtime API in 4 steps
Start with the call type that runs longest, since that is where token based billing grows fastest.
Step 1 of 4
List call types and costs
Pull a month of Realtime usage and divide audio token spend by call minutes for each call type, so the baseline is your own number.
Step 2 of 4
Rebuild the top call type
Recreate the highest volume call type on one finalist and run it next to Realtime on live traffic.
Step 3 of 4
Connect a phone number
On PyAI or Retell AI, assign a managed number. On Gemini Live, Deepgram or Realtime, route a Twilio or SIP trunk to the agent.
Step 4 of 4
Shift traffic gradually
Send a small share of calls first, compare reply delay, false interruptions and cost per minute, then raise the share.
Which OpenAI Realtime API alternative to pick
Pick by how much of the stack you want to keep. Gemini Live keeps the Realtime shape on Google models, Vapi and Deepgram let developers choose each vendor, and LiveKit gives you the source. For a phone agent with one predictable bill, PyAI Omni costs $0.05 a minute plus $0.01 for the number, about $7,200 a year at 10,000 minutes a month, while Retell AI suits teams that want a visual builder. A PyAI sandbox key costs nothing and opens with $50 in credit, no credit card needed.
More voice agent comparisons
OpenAI Realtime API alternatives FAQs
Common questions from teams weighing the Realtime API.
What does the OpenAI Realtime API cost per minute?
gpt-realtime bills $32 per million audio input tokens and $64 per million output. gpt-realtime-mini costs $10 and $20. Third party estimates convert the flagship to about $0.06 to $0.11 a minute, and AssemblyAI reports long calls running 2 to 5 times the headline because each turn resends context. No phone carrier is included.
What is a good OpenAI Realtime API alternative?
For phone agents, PyAI Omni at $0.05 a minute plus $0.01 for the number. Retell AI suits teams wanting a visual builder at $0.07, and ElevenLabs Agents suits voice first brands at $0.08 plus the LLM. Google Gemini Live is the nearest native audio model to gpt-realtime, while LiveKit Agents is the Apache 2.0 route for owning the code.
Does the OpenAI Realtime API support phone calls?
No, not on its own. The OpenAI Realtime API streams audio over WebRTC or WebSocket and offers no phone numbers, SIP trunks or call routing, so teams connect Twilio, LiveKit SIP or another carrier themselves. PyAI Omni, Retell AI and LiveKit Cloud sell numbers directly, so one account covers the whole call.
Why do teams leave the OpenAI Realtime API?
Cost, telephony and control. Audio tokens make long calls cost more per minute, there is no phone line, and one model runs speech, reasoning and voice, so no single layer can be swapped or tuned. AssemblyAI and Dasha also report false barge in on background noise and more than 30 event types to maintain.
Is there an open source OpenAI Realtime alternative?
Yes. LiveKit Agents, licensed Apache 2.0, and Pipecat, licensed BSD-2, are free frameworks you host yourself, and an open speech-to-speech model and an open multimodal model are open speech to speech models. Each needs servers, engineering time and paid model providers in return for full control. LiveKit Cloud and Pipecat Cloud offer hosted versions billed per minute.
How do I migrate off the OpenAI Realtime API?
Pick a managed engine or a framework. List call types with their current cost per minute, rebuild the highest volume one on a finalist, and compare reply delay, barge in and cost on live calls before moving more traffic. PyAI accepts OpenAI compatible requests and includes phone numbers, so a Realtime agent moves with modest client changes.
Test PyAI on your own audio and calls.
Get a sandbox key and hear the output in under a minute. Sandbox keys are free, with $50 in starter credit and no credit card required. Speak is $0.04/min, Cast $0.02/min, Hear $0.001/min and Omni $0.05/min + $0.01 telephony.