Every AI voice agent runs the same four-stage pipeline: speech-to-text turns the caller's words into tokens, an LLM decides what to say, text-to-speech turns that back into audio, and a telephony layer carries it over a phone line. The platforms below all orchestrate that loop in real time. What separates them is the price per minute, how fast they answer, and whether you can bring your own LLM or you are locked to theirs.
This compares Vapi, Retell AI, Bland AI, Synthflow, and ElevenLabs Agents on those three axes, with numbers taken from each vendor's own pricing and docs as of July 2026. Advertised latency figures are labeled as advertised, because none of these are audited SLAs.
Quick verdict
| If you are... | Pick | Why |
|---|---|---|
| A developer assembling your own stack | Vapi | $0.05/min orchestration, true OpenAI-compatible custom LLM, widest speech-provider list |
| Running high-volume phone operations | Retell AI | Batch and outbound campaigns, SIP to any carrier, SOC 2 Type 1 and 2 plus HIPAA built in |
| A regulated team wanting one closed vendor | Bland AI | In-house speech and model stack, data never routed to third parties, tiered per-minute pricing |
| A non-engineer building without code | Synthflow | No-code visual builder, hosted models, from $0.09/min voice engine |
| Prioritizing voice quality above all | ElevenLabs Agents | Best-in-class text-to-speech, ~75ms synthesis, custom LLM via server integration |
Before comparing features, decide whether you want to own the LLM. If yes, you are choosing between Vapi (paste an OpenAI-compatible URL), Retell (host a WebSocket relay), and ElevenLabs (server integration). If no, Bland gives you one vendor and a closed pipeline. That choice sets your latency floor, your cost curve, and where your call data lives, before you look at anything else.
How to actually compare voice platforms
Feature grids on vendor sites list dozens of integrations and hide the three things that decide whether an agent is usable and affordable. Compare on these:
- Price per minute, fully loaded. A low platform fee means nothing if the LLM and telephony are billed on top. Vapi's $0.05/min is orchestration only; Bland's $0.11/min bundles the model but not the phone line. Add every line before comparing.
- Response latency. A caller notices a pause. Under ~500ms feels like a conversation; past ~1000ms feels broken. The LLM is 350 to 1000ms of that budget, so the model you run is the single biggest lever.
- Bring-your-own-LLM. If the platform lets you point at your own model endpoint, you control cost, latency, fine-tuning, and data residency. If it does not, you inherit whatever the vendor picked.
The rest of this page is those three tables, then a per-platform breakdown, then how to pick.
Pricing, per minute
Voice pricing splits into a platform or voice-engine fee, an LLM cost, and telephony. Some vendors bundle, some pass through. All numbers are from each vendor's pricing page as of July 2026.
| Platform | Platform / voice fee | LLM | Effective range |
|---|---|---|---|
| Vapi | $0.05/min | Passthrough at cost, $0 with your own key | $0.05/min + |
| Retell AI | $0.055/min infra + $0.015 TTS | Passthrough $0.003-$0.345/min | $0.07-$0.31/min |
| Bland AI | Tiered: $0.14 free / $0.12 Build / $0.11 Scale | In-house, bundled | $0.11-$0.14/min + telephony |
| Synthflow | $0.09/min voice engine | $0.02 (mini) - $0.04/min (GPT-5.x) | ~$0.11-$0.16/min |
| ElevenLabs Agents | $0.08/min beyond plan minutes | Custom, billed separately | $0.08/min + LLM |
Vapi is cheapest on paper because the $0.05/min is orchestration only; the true cost depends on the speech and model providers you plug in, and drops toward that floor when you supply your own provider keys. Retell publishes the widest range because the LLM line swings from $0.003/min (a nano model) to $0.345/min (a realtime model). Bland's tiers require a $299/mo (Build) or $499/mo (Scale) subscription to reach the lower per-minute rate, and telephony is separate. Synthflow and ElevenLabs both bill the voice engine and the LLM as distinct lines.
At scale the LLM line dominates, not the platform fee. A $0.05/min orchestration fee is fixed; the model can be anywhere from $0.003 to $0.345 per minute depending on which one you run and how many tokens each turn consumes. Choosing and hosting the right model is the lever that moves the bill, which is the whole argument for bring-your-own-LLM below.
Response latency
Latency is the gap between the caller finishing their sentence and the agent starting to speak. Under ~500ms feels natural, up to ~800ms is acceptable, and past ~1000ms callers read the delay as a system failure. That budget is spent across the pipeline:
| Stage | Typical latency | Notes |
|---|---|---|
| Speech-to-text | 100-300ms | Deepgram advertises ~150ms |
| LLM (time-to-first-token) | 350-1000ms | Largest and most variable slice |
| Text-to-speech | 90-200ms | ElevenLabs advertises ~75ms synthesis |
| Network round trips | 50-200ms | Between vendors in a stitched pipeline |
The LLM is the biggest chunk and the one that varies most between models. A practical design target cited across voice-infra writeups is an LLM time-to-first-token around 500ms; keep the model under that and the full pipeline lands inside the ~800ms ceiling. Advertised platform numbers (not audited): Bland ~400ms (its product page claims sub-200ms), Vapi ~500 to 800ms, Retell ~600ms. Synthflow and ElevenLabs do not advertise a full voice-to-voice figure.
The takeaway is not which vendor's marketing number is lowest. It is that the model you run decides most of the budget, so on any platform that lets you bring your own LLM, a fast, low-time-to-first-token model is how you get under ~800ms and stay there. That is covered in the voice agent API and latency breakdown.
Bring your own LLM: who is open, who is closed
This is the axis that reorganizes the market. If a platform lets you point at your own model endpoint, you control the cost, the latency, whether you can fine-tune, and whether call data leaves your infrastructure. Support ranges from a clean base-URL swap to no support at all.
| Platform | Custom LLM? | Mechanism |
|---|---|---|
| Vapi | Yes | Set provider to custom-llm and paste any OpenAI-compatible base URL. Any model you host works unchanged. |
| Retell AI | Yes | Host a WebSocket server that speaks Retell's protocol and calls your model behind it. Retell recommends against it unless you have a reason. |
| ElevenLabs Agents | Yes | Connect a custom model through a server integration; the LLM is billed separately. |
| Synthflow | Partial | Choose from a set of hosted models (GPT family). Not an arbitrary endpoint. |
| Bland AI | No | API accepts only proprietary base and turbo models. No base URL, no external model. |
Vapi is the only true base-URL swap: its docs name OpenRouter, Together, and your own server as valid backends, so the endpoint just needs a streaming /chat/completions route. Retell supports a custom model but makes you host a relay, which is more work and adds a network hop. Bland is deliberately closed, which is the point of its pitch: one vendor, in-house models, and call data that never routes to a third party. The trade-off is you cannot swap in a faster or cheaper model when one ships.
The five platforms in detail
Vapi
Developer-first orchestration API. You assemble speech-to-text, LLM, and text-to-speech; Vapi runs the real-time loop and the telephony. $0.05/min platform fee, speech and model passthrough at cost. Up to 10 free US phone numbers, Twilio import for international, SIP trunking. Supports the widest speech-provider list (Deepgram, AssemblyAI, Cartesia for STT; ElevenLabs, Cartesia, PlayHT, Rime, and more for TTS). HIPAA is an opt-in add-on. Raised a $50M Series B in May 2026 ($72M total), and reports over a billion calls handled.
- Cheapest platform fee at $0.05/min
- True OpenAI-compatible custom LLM (paste a base URL)
- Widest choice of speech and voice providers
- Free US numbers plus SIP and Twilio
- You assemble and tune the stack yourself
- HIPAA is a paid add-on, not default
- Fully-loaded cost depends on the providers you pick
Retell AI
Built for running phone operations at volume. $0.07 to $0.31/min all-in, from $0.055/min of voice infrastructure plus a passthrough LLM line. Inbound, outbound, batch, and campaign tooling; built-in numbers at $2/mo or bring any SIP carrier (Twilio, Telnyx, Vonage). Speech and voice are bundled and metered, so there is no separate key or invoice. SOC 2 Type 1 and Type 2, HIPAA, and GDPR are documented, with a self-serve BAA. YC W24; reported around $60M ARR annualized as of April 2026.
- Batch and outbound campaign tooling built in
- SOC 2 Type 1 and 2, HIPAA, GDPR, self-serve BAA
- SIP to any carrier, built-in numbers at $2/mo
- Bundled speech and voice, single invoice
- Custom LLM needs a self-hosted WebSocket relay
- All-in price varies widely with the model chosen
- Less low-level control than assembling your own stack
Bland AI
Self-hosted, security-first, closed stack. Its own ASR, its own LLMs (base and turbo), and its own text-to-speech run on its own infrastructure, so call data never routes to a third party. Tiered pricing: $0.14/min free, $0.12/min on the $299/mo Build plan, $0.11/min on the $499/mo Scale plan, with telephony billed separately (bring your own carrier or use its managed Twilio). Known for outbound at scale. SOC 2 Type II, HIPAA, GDPR, and PCI DSS are listed on its site. Raised a $50M Series C in June 2026 ($100M+ total), and reports over 3.5M calls per week.
- One vendor, closed pipeline, data stays in-house
- Advertised ~400ms latency (product page claims sub-200ms)
- Strong compliance posture for regulated industries
- Proven at outbound scale
- No custom LLM; locked to base and turbo
- Lower per-minute rate requires a paid subscription
- Cannot adopt a faster or cheaper model when one ships
Synthflow
No-code visual builder aimed at non-engineers and agencies. Pay-as-you-go starts at $0.09/min for the voice engine plus the LLM ($0.02/min for a mini model, $0.04/min for GPT-5.x); a common configuration lands around $0.16/min with managed Twilio, or $0.11/min with your own Twilio and a mini model. Native telephony and SIP trunking. SOC 2, HIPAA, GDPR, and ISO 27001 badges. Enterprise tier starts around $30k/year. You choose from a set of hosted models rather than pointing at your own endpoint.
- No-code builder, fastest path for non-engineers
- Transparent per-component pricing from $0.09/min
- SOC 2, HIPAA, GDPR, ISO 27001
- Bring your own Twilio to cut cost
- Model choice limited to hosted options
- Less control than a developer API
- Enterprise features gated behind a ~$30k/year floor
ElevenLabs Agents
The agents layer on top of the best-known text-to-speech engine, so voice quality is the draw. Plans run from a free tier (15 minutes) up through Business ($990/mo, 12,375 minutes), with additional minutes at $0.08/min and the LLM billed separately. A custom LLM connects through a server integration, and you can select from supported third-party models. Text-to-speech synthesis is advertised at around 75ms. BAAs for HIPAA are available on the Enterprise tier.
- Best-in-class voice quality and voice library
- ~75ms text-to-speech synthesis
- Custom LLM via server integration
- Generous plan minutes at low tiers
- Per-minute overage plus separate LLM billing adds up
- Agents layer is newer than dedicated call platforms
- HIPAA BAA gated to Enterprise
If you want to own the whole pipeline, LiveKit and Pipecat are open-source frameworks for building real-time voice agents on your own infrastructure. You wire up the speech and model providers yourself and self-host, which is the most control and the most work. Both are the natural home for a self-hosted, low-latency model endpoint, covered next.
The model behind the agent
On every platform that supports a custom LLM (Vapi, Retell, ElevenLabs, and any self-hosted LiveKit or Pipecat stack), the model is the part you control, and it is the part that decides most of the latency budget and most of the per-minute bill. The lever is time-to-first-token: how fast the model starts streaming its answer after the caller stops talking.
Morph serves Gemma 4 31B (morph-gemma4-31b) behind one OpenAI-compatible API at api.morphllm.com/v1, so it drops into any of these platforms as a custom LLM by base URL. It is a dense 31B multimodal model with a 175k context window at $0.204 input / $0.548 output per million tokens. For voice specifically, a dedicated single-tenant endpoint is the recommended setup: it delivers sub-300ms time-to-first-token with no shared-tenant jitter, which keeps the LLM comfortably inside the ~800ms conversation budget. The full pipeline argument, and how to wire Morph into Vapi or Retell, is in the voice agent API and latency guide.
Which to pick
- Vapi if you are a developer who wants control and the cheapest orchestration fee, and you are comfortable assembling the stack and bringing your own model.
- Retell AI if you are running phone operations at volume and want campaigns, SIP, and compliance without building them.
- Bland AI if you are in a regulated industry and value one closed vendor with in-house models over the ability to swap the LLM.
- Synthflow if you are a non-engineer or agency who needs a working agent from a visual builder.
- ElevenLabs Agents if voice quality is the deciding factor and you want the best text-to-speech under the agent.
For the head-to-head feature and pricing matrix across the four developer platforms, see Vapi vs Retell vs Bland vs Synthflow. For the model and latency layer underneath all of them, see the voice agent API guide.
Frequently asked questions
What is the best AI voice agent platform in 2026?
It comes down to whether you want to control the LLM. Vapi for developers who assemble their own stack at $0.05/min orchestration; Retell for high-volume production call operations at $0.07 to $0.31/min; Bland for regulated teams that want a closed in-house stack at $0.11 to $0.14/min; Synthflow for non-engineers building without code; ElevenLabs Agents when voice quality is the priority.
How much does an AI voice agent cost per minute?
Vapi is $0.05/min plus speech, LLM, and voice at cost. Retell is $0.07 to $0.31/min all-in. Bland is $0.11 to $0.14/min tiered, telephony separate. Synthflow starts at $0.09/min for the voice engine plus $0.02 to $0.04/min for the model. ElevenLabs Agents is $0.08/min beyond plan minutes plus a separate LLM line.
Which voice AI platforms let you use your own LLM?
Vapi (paste an OpenAI-compatible base URL), Retell (host a WebSocket relay), and ElevenLabs (server integration) all support a custom LLM. Bland does not; it accepts only its base and turbo models. Synthflow lets you pick from hosted models rather than an arbitrary endpoint.
What latency should a voice AI agent have?
Under ~500ms feels natural, up to ~800ms is acceptable, past ~1000ms feels broken. The LLM is 350 to 1000ms of that budget, so the model's time-to-first-token is the lever that matters most. Advertised: Bland ~400ms, Vapi ~500 to 800ms, Retell ~600ms.
Is Vapi or Retell better?
Vapi is the developer-first orchestration layer with the cheapest platform fee and a true bring-your-own-LLM base URL. Retell is built for phone operations at volume with campaigns, SIP, and compliance out of the box. Build on Vapi; operate on Retell.
Related
Run the model behind your voice agent on Morph
Vapi, Retell, and ElevenLabs all let you point the agent at your own LLM. Morph serves Gemma 4 31B on an OpenAI-compatible API, with dedicated endpoints tuned for sub-300ms time-to-first-token.