Vapi
Composable voice AI: bring your own transcription, model, and voice provider
Vapi is a developer platform for building voice AI agents that orchestrates best-of-breed components rather than supplying its own. You choose the speech recognition provider, the language model, and the text-to-speech voice, and Vapi handles the real-time orchestration, turn taking, interruption handling, telephony, and tooling that makes them behave like a conversation.
Overview
The hard part of a voice agent is not any single component; it is the orchestration between them. Deciding when the caller has finished speaking, cutting off generation when they interrupt, filling silence naturally, recovering when transcription is wrong, and doing all of it inside the second of latency a conversation allows. Vapi's product is that orchestration layer, deliberately unopinionated about which providers sit underneath.
That composability is its defining choice. A team can run one transcription provider for accuracy in a specific language, a fast model for routine turns and a stronger one for complex reasoning, and whichever voice sounds right for the brand, changing any of them without rebuilding. In a field where the underlying models improve every few months, not being locked to one vendor's stack is a genuine strategic advantage.
It is squarely a developer platform, priced per minute on top of the underlying provider costs, with an SDK and API as the primary interface and a dashboard for configuration and testing. Teams wanting a finished receptionist product will find it a construction kit rather than an answer. Teams building voice into a product, especially those who care about model choice, will find it the most flexible option in the category.
Best for
Engineering teams building voice agents into products or operations who want control over the model and voice stack, and technically capable agencies deploying customized voice solutions.
Not the right fit for
- Non-technical buyers wanting a configured phone receptionist without development.
- Teams that would rather not make architectural decisions about transcription and model providers.
- Organizations needing a full contact center platform with agent desktops and workforce management.
- Regulated scripts requiring guaranteed verbatim delivery without substantial validation work.
- Businesses in markets where automated calling is heavily restricted.
How it works
- 1
You define an assistant: a system prompt describing its role, the transcription provider, the language model, the voice, and configuration for turn taking, interruption sensitivity, and filler behavior.
- 2
Tools are attached so the agent can act during a conversation, calling your APIs to check availability, retrieve records, or trigger workflows, with structured parameters extracted from what the caller said.
- 3
The assistant is connected to telephony for inbound and outbound calls, or embedded in a web or mobile application through the SDK for in-product voice.
- 4
During the call Vapi orchestrates the pipeline in real time, streaming audio to transcription, prompting the model, and synthesizing speech while managing interruptions and endpointing. Afterwards it delivers transcripts, recordings, structured data, and analysis through webhooks and the API.
Feature breakdown
20 features in 4 modulesComposable pipeline
The architecture that distinguishes it.- Provider choice at every stage
- Select transcription, language model, and voice independently from supported providers and change any of them without rebuilding.
- Model routing
- Use a fast model for routine turns and escalate to a stronger one when the conversation needs it, balancing latency against capability.
- Bring your own keys
- Run on your own provider accounts so model and voice costs are billed directly and negotiated independently.
- Custom voice support
- Cloned and branded voices from supported vendors rather than a fixed in-house library.
- Custom LLM endpoints
- Point the assistant at your own hosted model for teams with fine-tuned or self-hosted requirements.
Conversation orchestration
The real-time work that makes it feel like talking.- Endpointing and turn taking
- Deciding when the caller has finished speaking, tuned per assistant, which is the single biggest determinant of whether a call feels natural.
- Interruption handling
- Generation stops when the caller speaks, with configurable sensitivity to avoid cutting off on background noise.
- Filler and backchannel behavior
- Natural acknowledgements and pauses that mask processing time rather than leaving dead air.
- Background noise handling
- Filtering and configuration for real-world call conditions rather than studio audio.
- Latency optimization
- Streaming across every stage so the first audio arrives as early as possible in the response.
Actions and telephony
Doing work and reaching the phone network.- Tool calling
- The agent calls your APIs mid-conversation with structured parameters extracted from what the caller said.
- Inbound and outbound calling
- Numbers provisioned in-platform or brought from an existing telephony provider, with programmatic dialing.
- Call transfer
- Escalation to humans with context, plus routing rules for when an agent should not continue.
- Web and mobile SDKs
- In-product voice experiences using the same assistants that handle phone calls.
- Squads and multi-agent flows
- Handoff between specialized assistants within a call rather than one agent attempting everything.
Operations
Running it in production.- Call logs and recordings
- Full transcripts, audio, and per-turn timing for debugging why a conversation went wrong.
- Structured data extraction
- Post-call extraction of fields and outcomes for delivery to CRM and workflow systems.
- Testing and evaluation
- Simulated conversations and evaluation runs to catch regressions before customers hear them.
- Concurrency controls
- Limits and scaling arrangements for campaign volume and peak inbound load.
- Webhooks and server events
- Real-time events during calls plus post-call delivery for integration with your own systems.
Use cases
4 documentedSaaS product adding a voice interface
Users want to interact by speech in the mobile app, and the team needs the same logic to serve phone calls.
One assistant definition powers both the in-app SDK and the phone line, with tools calling the product's existing API.
Startup optimizing for multilingual accuracy
Transcription quality varies by language and no single provider is best across the markets served.
Provider selection per assistant lets each market use the transcription and voice that perform best there.
Operations team automating routine inbound calls
A high share of calls are status checks and simple bookings that consume staff time.
An agent handles those with tool calls into the operational system, transferring anything unusual to a person.
Agency building bespoke voice solutions
Clients need customized agents rather than a templated receptionist, with different stacks per use case.
Composability allows a different pipeline per client without changing platforms or rebuilding orchestration.
Pricing
from From roughly $0.05 per minute platform fee, plus provider costsUsage-based platform fee per minute on top of the underlying transcription, model, and voice provider costs, which can be billed through Vapi or directly on your own provider accounts. Enterprise arrangements for volume and dedicated capacity.
| Plan | Price | Includes |
|---|---|---|
| Pay as you go | From about $0.05 per minute plus provider costs |
|
| Scale | Negotiated rates committed volume |
|
| Enterprise | Quoted annual |
|
Billing notes
- The headline per-minute rate is the platform fee only; transcription, model, and voice provider costs are additional and often exceed it.
- Bringing your own provider keys makes costs transparent and negotiable but shifts vendor management to you.
- Model choice materially affects cost per minute, so a fast cheap model for routine turns is an economic decision as well as a latency one.
- Telephony rates vary by destination and are billed separately from platform and provider costs.
- Rates as published August 2026; per-minute pricing across this category has fallen consistently and should be reconfirmed.
Value assessment: Vapi's value is optionality. The underlying models change faster than any product roadmap, and a platform that lets you swap a transcription provider or move to a better voice without rebuilding protects against being locked to whichever vendor was best in the quarter you started. The cost of that flexibility is decisions: someone has to choose and maintain the stack, and the true per-minute price is less legible than an all-inclusive rate. For teams that want control, the trade is worth it; for teams that want an answer, it is friction.
Strengths & limitations
Strengths
- Provider independence at every stage, which protects against lock-in in a fast-moving model landscape.
- Strong orchestration around endpointing, interruption, and latency, which is where voice agents actually succeed or fail.
- Tool calling and custom model endpoints make sophisticated integrations possible.
- Web and mobile SDKs extend the same assistants beyond the phone into products.
- Detailed call logs with per-turn timing, which is essential for debugging conversational failures.
- Transparent usage pricing with the option to run on your own provider accounts.
Limitations
- Developer platform with no realistic non-technical path.
- True cost is spread across multiple vendors, making budgeting less straightforward.
- Provider choice means provider maintenance, including monitoring quality changes upstream.
- Quality depends heavily on configuration and prompt design rather than working well by default.
- Automated calling and AI disclosure rules vary by jurisdiction and are tightening.
- Not a contact center platform, so workforce and queue management need other systems.
Head-to-head comparisons
5 alternativesVapi vs Bland AI
from From roughly $0.09 per minute of call time, with free credits to startThe central comparison in developer voice AI. Bland runs its own integrated stack and argues for latency and consistency; Vapi orchestrates providers you choose and argues for flexibility and avoiding lock-in. Teams with strong opinions about models and voices, or with multilingual requirements, generally prefer Vapi. Teams who want fewer decisions and one vendor prefer Bland.
Full Vapi vs Bland AI comparisonVapi vs Retell AI
from From roughly $0.07 per minute combined, with free credits to startClose competitors serving the same developer audience with similar per-minute economics. Retell tends to be praised for a smoother out-of-the-box experience and flow building, while Vapi goes further on composability and provider control. The practical way to choose is to build the same agent on both and compare recordings under realistic conditions.
Full Vapi vs Retell AI comparisonVapi vs Synthflow
from From roughly $29 per month with bundled minutesOpposite ends of the same market. Synthflow packages voice agents for non-technical buyers with templates, a visual builder, and agency-friendly white labeling. Vapi is infrastructure requiring engineering. Businesses wanting a working receptionist next week should look at Synthflow; teams building voice into a product should look at Vapi.
Full Vapi vs Synthflow comparisonVapi vs Millis AI
from From roughly $0.02 to $0.05 per minute depending on configuration, plus telephonyBoth are per-minute developer platforms, but Vapi's value is composability: transcription, model, and voice chosen independently, with bring-your-own-keys billing. Millis ships a fixed pipeline tuned for sub-second responses at roughly $0.02 per minute. Teams that want architectural choice accept Vapi's platform fee on top of provider costs; teams optimizing purely for cost per minute at volume will find Millis materially cheaper.
Full Vapi vs Millis AI comparisonVapi vs PlayAI
from Subscription tiers from low monthly amounts plus usage; conversational agents priced per minutePlayAI builds its own speech models and sells the agent layer on top, so voice quality, cloning, and delivery control come from one vendor. Vapi owns no models and instead lets you attach whichever synthesis provider sounds right. Products where an expressive branded voice is the differentiator suit PlayAI; teams that want to swap voice and model vendors as the market moves suit Vapi.
Full Vapi vs PlayAI comparisonImplementation & onboarding
- Setup time
- A prototype in an afternoon. A production agent takes weeks, most of it spent tuning endpointing, testing against realistic audio conditions, and designing escalation.
- Learning curve
- Moderate to steep. The API is approachable, but choosing providers, tuning turn taking, and designing conversations that survive unexpected input require iteration and real listening to failed calls.
- Onboarding
- Self-serve with documentation, examples, and an active developer community, with dedicated support on larger arrangements.
- Migration notes
- Assistant configurations and prompts are platform-specific and do not transfer. Because Vapi supports bringing your own provider keys, the underlying model and voice relationships can survive a platform change, which is a genuine reduction in switching cost compared with integrated stacks.
Platform, API & security
- Platforms
- REST APIWeb and mobile SDKsTelephony integrationWeb dashboard
- API
- Comprehensive API and SDKs for assistants, calls, tools, and events, with real-time server events during calls, webhook delivery, and support for custom model endpoints.
- Compliance
- GDPRCCPASOC 2HIPAA support on qualifying arrangementsTCPA considerations for outbound
- Data residency
- Regional options depending on arrangement and provider selection.
- SSO
- Available on enterprise plans.
- Security notes
- Because the pipeline spans multiple providers, the data protection assessment covers each one; bringing your own keys concentrates the relationship with vendors you have already reviewed, which is often the cleanest position for regulated deployments.
Support & resources
- Channels
- Documentation and developer communityEmail supportDedicated support on enterprise plans
- Documentation
- Detailed developer documentation covering assistants, tools, telephony, SDKs, and configuration of turn taking and latency behavior.
- Community
- One of the more active developer communities in voice AI, with substantial shared knowledge on provider selection and conversation tuning.
Company
- Founded
- 2023
- Headquarters
- San Francisco, California, United States
- Ownership
- Private, venture-backed
- Employees
- ~50 (est. 2026)
- Funding
- Raised venture funding including a Series A round.
Timeline
- 2023Founded as an orchestration layer for voice AI, deliberately provider-agnostic.
- 2024Adds tool calling, SDKs, and multi-agent flows as production deployments grow.
- 2025Raises venture funding as voice agents move from experiment to operational deployment.
- 2026Established as the composable option in developer voice AI, favored by teams wanting provider control.
Integrations
- Twilio
- Deepgram
- ElevenLabs
- OpenAI
- Anthropic
- Cartesia
- Make
- n8n
- Zapier
Frequently asked questions
10 questionsWhat is Vapi?
Vapi is a developer platform for building voice AI agents. It orchestrates the real-time pipeline, transcription, language model, and speech synthesis, while letting you choose which provider handles each stage, and manages telephony, turn taking, interruption handling, and tool calling around them.
How much does Vapi cost?
There is a platform fee from around $0.05 per minute, on top of the underlying transcription, model, and voice provider costs, plus telephony rates by destination. Provider costs can be billed through Vapi or directly on your own accounts, and often exceed the platform fee itself.
Vapi vs Bland AI: what is the real difference?
Architecture. Bland runs its own integrated speech and model stack, arguing for latency and consistency. Vapi orchestrates providers you select, arguing for flexibility and protection against lock-in as models improve. Choose Vapi if you want control over the stack, Bland if you want fewer decisions.
Why does provider choice matter?
Because the underlying models change constantly. Transcription accuracy varies by language and accent, voice quality varies by vendor, and model capability and cost shift every few months. Being able to swap any stage without rebuilding the orchestration is a genuine advantage over a fixed stack.
What is endpointing and why does it matter so much?
Endpointing is deciding when the caller has finished speaking. Cut in too early and you interrupt them, wait too long and the conversation feels sluggish and robotic. It is the single most consequential setting in a voice agent, and tuning it against realistic audio is most of what separates a good deployment from a bad one.
Can Vapi agents do things, not just talk?
Yes, through tool calling: the agent invokes your APIs during the conversation with structured parameters extracted from speech, so it can check availability, look up an account, or create a record. That capability is what makes an agent useful rather than merely a talking form.
Can I use it inside my app rather than on the phone?
Yes. Web and mobile SDKs let the same assistant definitions power in-product voice experiences alongside phone calls, which is useful when the same logic should serve both channels without maintaining two implementations.
Do I need to disclose that callers are speaking to an AI?
Increasingly yes, with disclosure requirements introduced or under consideration in multiple jurisdictions and expectations tightening generally. Beyond compliance, undisclosed automation damages trust badly when discovered, so building disclosure into the opening line is the sensible default.
Is Vapi suitable for a non-technical business?
Not really. It is infrastructure, and getting good results requires engineering work in configuration, integration, and iterative testing. Businesses wanting a working receptionist configured through a visual builder should look at no-code voice platforms instead.
How reliable are voice agents in real conditions?
Solid for bounded, structured conversations such as booking, confirmation, and qualification; still fragile in open-ended ones, with strong accents, poor line quality, and unexpected turns as common failure modes. Evaluate with recordings of real calls rather than scripted demonstrations before committing to customer-facing deployment.
Editorial verdict
Vapi bet that the components of voice AI would keep improving faster than any one vendor could keep up, and that orchestration, not the models, was the durable product. That bet looks correct. Being able to change transcription providers for a new market, swap in a better voice, or route between models for cost and latency without rebuilding the conversation layer is worth real money over the life of a deployment. What it asks in return is engineering judgment: choose the stack, tune the endpointing, test against messy audio, and design the escalation path. Teams that want to own those decisions get the most flexible platform in the category. Teams that want the decisions made for them should look elsewhere and will be happier for it.
Written by the SaaSTracker editorial team. Awards, when shown, are judged against the published criteria in our methodology.