AI voice agents can listen, generate responses, and speak in a call or application. This comparison includes finished platforms, configurable infrastructure, and raw APIs; model access alone is not a deployed telephone-answering service.
Quick answer
- Pay-as-you-go platform: Retell advertises US$0.07–US$0.31/minute for supported standard configurations; speech-to-speech options and complete costs depend on architecture, model, telephony, and add-ons.
- Bundled core processing: Bland Start costs US$0.14/minute with LLM, speech recognition, and synthesis included. Transfers and other applicable services are separate.
- Enterprise implementation: PolyAI uses custom pricing. Confirm language coverage, service levels, integrations, and contractual safeguards for the intended deployment.
OpenAI advertises GPT-Live-1 voice sessions at US$0.05/minute, billed per second, but backend models, tools, telephony, infrastructure, and development can add costs. This is an API building block, not an equivalent finished-phone-agent price.
Evaluation methodology
This comparison uses official pricing, product documentation, vendor-published benchmarks, and relevant external research. It does not represent standardized independent call-quality or latency testing across all twelve entries.
Criteria include usable cost, architecture, benchmark methodology, concurrency, telephony, CRM and calendar integration, language coverage, escalation, and deployment requirements. Published latency and performance results are attributed; differing definitions are not treated as directly comparable.
Dollar prices are US dollars. Base platform or model charges may exclude carrier fees, model usage, synthesis, recognition, recording, compliance add-ons, support, and implementation. Quotes should identify each component. An inaccessible page does not establish that a product is unavailable.
The 12 tools at a glance
| Tool | Product / use case | Cost basis | Advertised starting price | Trial / procurement |
|---|---|---|---|---|
| Retell AI | Configurable voice-agent platform | Architecture, models, telephony and add-ons affect total | US$0.07–US$0.31/min standard configurations | Self-serve; confirm credits |
| Vapi | Developer voice-agent infrastructure | Hosting plus STT, LLM, TTS and telephony | Approximately US$0.05/min hosting | US$5 credits advertised |
| Bland AI | Outbound and inbound voice workflows | Core STT, LLM and TTS bundled; transfers extra | US$0.14/min Start | Confirm current trial terms |
| ElevenLabs Agents | Voice-agent platform | Subscription allowances; external LLM and telephony extra | US$0.08/min additional minutes | Plan-specific free allowance |
| Twilio Conversation Relay | Conversational orchestration | Voice, telephony and selected model extra | US$0.07/min orchestration | Confirm trial credits |
| OpenAI Voice Agents API | Raw API building block | Backend tools, models, carrier and development extra | GPT-Live-1 US$0.05/min, billed per second | API billing |
| Google Gemini Live | Live audio API building block | Model-specific input/output rates; deployment extra | Gemini 3.8 Live / 3.1 Flash Live Preview: US$0.005 input + US$0.018 output/min | Model-specific availability |
| PolyAI | Enterprise voice-agent implementation | Custom deployment and contractual scope | Quote | Demo / negotiated pilot |
| SoundHound AI | Industry voice and conversational systems | Product, integration and deployment scope | Quote | Demo |
| Regal | Customer service and booking workflows | Deployment-specific integrations and safeguards | Quote | Demo |
| Observe.AI | Contact-center agents and QA | Coverage claims describe review scope, not accuracy | Quote | Demo |
| Crescendo | Managed customer-experience platform | Vendor rollout targets require contractual confirmation | Quote | Demo |
Prices describe advertised components or configurations, not independently verified all-in deployment costs. Quote-based entries require direct confirmation.
The 12 tools, reviewed
Retell AI
Retell AI offers pay-as-you-go voice agents with multiple voice architectures and separately itemized components.
The pricing page advertises US$0.07–US$0.31/minute for supported standard configurations, with 20 concurrent calls included. Total cost depends on LLM, voice infrastructure, synthesis, carrier, architecture, and optional features. Speech-to-speech configurations are available but should be priced separately from cascaded configurations.
Retell publishes a vCX benchmark, which is vendor-run evidence rather than a standardized independent comparison. Optional knowledge-base, QA, privacy, and outbound features have separate allowances or charges to confirm.
The changelog is unusually informative, and two features stand out. Conductor simulates a scenario - “what happens if the caller wants to reschedule?” - against your agent without dialling a real number, and it turns failed production calls into automatic test cases; Retell says it already handles 70% of its own simulation testing internally. And Colloquial Model rewrites phrasing in about 50 milliseconds to add fillers and human cadence without touching your prompts. Retell also ships Concurrency Blast at $0.1/minute for spikes up to 3× your limit, and added iOS/Android call-screening handling so agents don’t hang up on a screened call.
The limitation: compare the full selected configuration, including models, carrier, knowledge-base usage, QA, privacy features, and concurrency charges. A per-minute entry rate alone does not determine monthly cost.
Retell itemizes voice infrastructure and synthesis; concurrency and additional features can introduce separate billing rules. Use the selected configuration in the pricing calculator rather than treating the entry rate as a bundled speech-to-speech quote.
Vapi
Vapi is a developer-first AI voice agent platform that passes model and telecom costs through at cost, so you pay roughly $0.05/min for hosting plus whatever the underlying providers charge.
Vapi charges approximately US$0.05 per hosted call minute, while recognition, language models, synthesis, and telephony may be billed separately. Its calculator estimates roughly US$82–US$129 for 1,000 minutes under selected configurations, before other applicable charges. This is an illustrative workload estimate, not a guaranteed price for every implementation.
Vapi advertises US$5 in credits and self-serve access. Confirm included numbers, concurrency, and optional support charges for the selected account.
The limitation: the deployment team must configure telephony, integrations, call flows, and applicable safeguards. Vapi advertises a US$2,000/month HIPAA add-on; confirm its scope and contractual requirements.
Vapi supports configurable providers. Confirm architecture, provider charges, integrations, escalation, consent handling, and operational responsibility before treating flexibility as lower total cost.
Bland AI
Bland AI is a bundled AI voice agent platform with one all-in per-minute rate - $0.14 on Start, $0.12 on Build - and no separate LLM, speech-to-text or text-to-speech charges.
Bland’s Start tier costs US$0.14/minute without a platform subscription fee; Build advertises US$0.12/minute plus US$299/month. Core LLM, STT, and TTS processing is bundled, while transfer time and additional services may be charged separately. Predictable component packaging does not establish the lowest complete price.
Bland leans into outbound and appointment scheduling - its Build/Enterprise tiers include an appointment scheduling node, SMS, and warm/live transfers. Bland’s docs claim sub-400ms latency and 40+ languages.
The limitation: compare equivalent architecture, transfer usage, telephony, concurrency, and subscription fees. Different advertised component floors are not directly comparable complete call costs.
ElevenLabs Agents
ElevenLabs Agents offers configurable voices and subscription minute allowances, with additional minutes advertised at US$0.08/minute.
ElevenLabs Agents includes plan-specific monthly call-minute allowances. Additional minutes cost US$0.08/minute, with higher burst rates when concurrency limits are exceeded. External model usage and carrier or telephony charges can apply separately. Its subscription and overage rate should not be treated as the complete cost of every deployment.
The limitation: unused subscription capacity, model charges, burst usage, and carrier costs can change the effective per-minute price. Evaluate voice quality using your actual scripts and callers.
Other platforms can integrate ElevenLabs voices. Confirm which provider, voice model, and billing arrangement the chosen configuration uses.
Twilio Conversation Relay
Twilio Conversation Relay provides conversational orchestration through Twilio Voice, with a separate component rate.
Conversation Relay pricing lists US$0.07/minute for orchestration. Associated voice, telephony, and selected language-model costs are additional. The API supports voice-processing and interruption workflows, but a complete deployment still needs the application and connected services.
Twilio reports p50 latency of 491ms and p95 of 713ms in its own benchmark. These are vendor-measured results under stated conditions, not independent testing or a directly comparable measure against every other vendor’s latency figure.
The limitation: bring-your-own-LLM means you’re still assembling the brain, and the pricing advantage shrinks if you route to an expensive model. Best if Twilio is already your comms backbone.
OpenAI Realtime and Voice Agents API
The OpenAI Voice Agents API is a raw speech-to-speech building block where GPT-Live-1 costs just $0.05 per minute, billed per second, rather than a finished agent.
OpenAI’s GPT-Live-1 documentation lists voice sessions at US$0.05/minute, billed per second without whole-minute rounding. Backend model and tool usage is separate. Add carrier charges, application infrastructure, orchestration, integration, and development when estimating a functioning phone-agent deployment. Realtime models have their own token-based rates; they are not interchangeable with this session price.
OpenAI also publishes a voice latency and cost guide that walks through the trade-offs - useful reading regardless of which vendor you pick.
The limitation: you own everything - telephony, telephony consent, orchestration, guardrails, escalation. Most teams underestimate that. Pick it when you already have an engineering team that wants a bespoke agent.
Google Gemini Live
Google Gemini Live provides live-audio models rather than a finished phone-agent service, with pricing that varies by model.
The Gemini API Standard pricing lists estimated audio input of US$0.005/minute and output of US$0.018/minute for Gemini 3.8 Live and Gemini 3.1 Flash Live Preview. Charges depend on the selected model and token consumption; telephony, tooling, application infrastructure, and integration are additional. Do not apply those rates indiscriminately across the Live model family.
The limitation: as with OpenAI, Gemini Live is a model, not an agent platform - telephony, compliance and orchestration are yours to build. And Google’s voice-agent tooling is younger than the OpenAI Realtime ecosystem.
Worth knowing before you commit: Gemini Live’s automatic language detection is genuinely useful for multilingual inbound, but the surrounding agent scaffolding - call routing, escalation, calendar and CRM writes, consent logging - all still has to be assembled. Budget for that engineering, not just the token bill.
PolyAI
PolyAI is an enterprise AI voice agent company that sells a custom-per-minute “Agentic Dialog Platform” - no public rate card, but a very credible operation.
PolyAI is the enterprise heavyweight on this list. It’s priced per minute on a custom quote that “includes proactive performance improvements, maintenance and 24/7 support.” What you buy is the stack: their Raven LLM, trained on 1B+ real enterprise conversations, plus a purpose-built Owl ASR tuned for accents and background noise, and a voice layer that blends human recordings with neural synthesis. They claim sub-second answers and strong business-rule adherence.
PolyAI advertises integrations with CRM and contact-center systems, multilingual agents, and contractual service levels. It supports HIPAA-related deployments with appropriate contractual and technical safeguards; HIPAA is not a general product certification. Confirm applicable security attestations, BAA, retention, and deployment scope directly.
PolyAI’s vendor-run Dialog-Eval reports a 280ms median for Dialog-RSN-1 versus 860ms for GPT-realtime-2.1 under its test conditions. A separate model scores responses without knowing system identity, but PolyAI designs and operates the benchmark. This is not independent third-party benchmark administration.
PolyAI reports selected customer containment and latency improvements. These are vendor-published outcomes, not independently replicated results or a forecast for another deployment.
PolyAI describes Dialog-RSN-1 as reasoning over raw audio while using a separately controlled synthesis layer. Roadmap descriptions should not be treated as generally available deployment choices; confirm the production architecture in the proposal.
The limitation: implementation timelines and responsibilities depend on the agreed integrations, languages, testing, and contractual scope. Obtain a deployment schedule rather than treating an estimated rollout as a guarantee.
SoundHound AI
SoundHound AI is a public voice-AI company (NASDAQ: SOUN) whose OASYS platform spans automotive, drive-thru and enterprise banking, and which just closed a major acquisition of LivePerson.
SoundHound is a publicly traded company with a long history in voice AI, spanning automotive, restaurant, and enterprise applications. Pricing is quote-based.
SoundHound announced completion of the LivePerson acquisition on September 4, 2026. Evaluate the current combined offering and integration roadmap against required workflows rather than assuming acquisition alone establishes product integration.
The limitation: confirm the available enterprise workflow, carrier and CRM integrations, implementation scope, and support commitments in a written quote.
Regal
Regal is an enterprise AI voice agent platform for regulated customer-experience teams, boasting a 97% containment rate claim, 40+ integrations and 30+ languages, priced on a quote.
Regal advertises containment as high as 97% and cost-to-serve reductions up to 80% in promotional material. These are vendor claims for selected conditions, not universal outcomes. It advertises qualification, appointment scheduling, reminders, and collection workflows; confirm the purchased scope.
Regal offers HIPAA-related safeguards and contractual support for eligible deployments. Confirm the BAA, data handling, and workflow-specific obligations rather than assuming a general compliance claim makes every use lawful.
The limitation: every headline number is Regal’s own marketing, not independently audited, and it’s a sales-led platform with no self-serve pricing. Budget a real evaluation cycle.
Observe.AI
Observe.AI is an enterprise agentic CX platform for large contact centers that unifies AI agents, agent-assist and conversation intelligence - priced on a quote.
Observe.AI combines customer-facing agents with contact-center evaluation and coaching. It advertises evaluation coverage of 100% of interactions; that describes coverage rather than guaranteed accuracy. Confirm workflow controls, human review, security attestations, and any eligible BAA separately.
The limitation: it’s an enterprise platform (often “a month or two” to production), not something you spin up in an afternoon, and there’s no public pricing. If you want an AI receptionist, this is overkill; if you run a 500-seat contact center, it’s the right shape.
Observe.AI advertises automated quality analysis across 100% of supported interactions. Coverage is different from scoring accuracy: human review, calibration, and auditing remain necessary. This does not establish that smaller QA products are disappearing.
Crescendo
Crescendo is an AI-native CX platform (CXP) that positions itself as replacing your chatbot, CCaaS, ticketing, QA, VoC and knowledge tools with a team of autonomous agents - priced on a quote.
Crescendo advertises consolidation of CX workflows with controls over agent actions and human approvals. Its approximately 30-day go-live is a vendor delivery target; actual timing depends on integrations, security review, workflow design, and customer readiness. Resolution and cost-reduction figures are vendor claims, not expected results for every customer.
The limitation: it’s a full CX-suite replacement story, so it’s a heavier commitment than picking a voice agent - and the market for “replace your whole stack” is crowded and expensive. Best suited to mid-market teams consolidating sprawl, not to a team that just needs after-hours call answering.
Which tool is actually best for appointment booking?
Many entries can support booking through native integrations, configurable actions, or custom development. This does not establish that all twelve provide a native production-ready booking service.
Retell Workflows, introduced in September 2026, integrates external Salesforce, HubSpot, Zendesk, Cal.com, and Calendly data before or after calls. It is an orchestration and integration layer, not a standalone CRM.
PolyAI and Regal advertise enterprise booking workflows, while Observe.AI and Crescendo place agent actions within broader CX processes. Raw APIs require calendar access, authentication, booking logic, and operational controls to be built.
Test availability, time zones, daylight-saving changes, ambiguous dates, duplicate prevention, slot reservation and confirmation, calendar failures, and human escalation. Confirm which actions and safeguards are available in the purchased configuration.
What to look for when you buy voice AI
The platforms differ less on “can it talk” and far more on the fine print. Before you sign anything, check these:
- What a minute actually costs end to end. Is the number bundled or a teaser? Bland’s $0.14 covers LLM+STT+TTS; Vapi’s ~$0.05 hosting is plus models. Retell’s $0.07–$0.31 range depends entirely on which voice and model you pick. Get the all-in figure, including telephony, before comparing.
- Concurrency, because it’s the real bill. Retell gives you 20 concurrent free then $8 each; Vapi’s Core is 10 lines and extra lines are $10/month; ElevenLabs scales 4→40 by tier and charges double for burst. Most surprise invoices are concurrency and overage, not the headline rate.
- Consistent latency definitions. Ask whether measurements cover first token, first audio, or end-to-end response, and request conditions and percentiles. An unpublished figure does not prove latency was never measured.
- Transfer and escalation. A voice agent that can’t hand off to a human with context is a dead end. Check warm transfer, and whether the human sees the transcript.
- Deployment-specific safeguards. Confirm consent and opt-outs, recording rules, retention, privacy controls, and eligible BAA terms. Vapi advertises a US$2,000/month HIPAA add-on; other vendors’ contractual support and costs need direct confirmation.
Evaluate latency and call quality together
First-token, first-audio, and end-to-end conversational response times are different measurements. Vendor results may use different hardware, networks, prompts, languages, and turn-detection assumptions. No universal 500ms threshold establishes whether a call feels human.
Twilio reports p50 491ms and p95 713ms in an internal benchmark. PolyAI’s vendor-designed Dialog-Eval reports median 280ms for Dialog-RSN-1 versus 860ms for GPT-realtime-2.1; a blinded model judge does not make its administration independent. Treat these as evidence about the stated tests rather than a universal vendor ranking.
Speech-to-speech systems process audio directly, while cascaded systems separate recognition, language reasoning, and synthesis. Retell supports multiple configurations. Audio-aware reasoning with separate synthesis is another design option. Architecture affects control and information flow, but it does not by itself prove better latency, interruption handling, or reliability.
Evaluate realistic phone audio, interruption, background noise, escalation, and complete task success alongside response timing.
The legal reality: consent, identification, and recording
The FCC’s AI-voice declaratory ruling establishes that AI-generated voices fall within the TCPA’s artificial-or-prerecorded-voice restrictions. It does not make every use of AI in a call automatically unlawful.
Covered telemarketing calls generally require appropriate prior consent, often prior express written consent. Requirements vary by purpose, recipient, technology, exemptions, and jurisdiction. Evaluate caller identification, opt-outs, calling hours, do-not-call restrictions, recording consent, privacy, and state requirements for the specific workflow.
AI-use disclosure at the start of a conversation can support transparency, but should not be presented as a universal standalone federal duty for every AI interaction. Identification rules and specific state disclosure laws require separate assessment.
Consent revocation and one-to-one consent are distinct legal issues. This comparison does not rely on an uncited characterization of a 2026 revocation change. Maintain appropriate consent records, suppression processes, and escalation controls, and obtain qualified review for the intended calling program before deployment.
Who should skip voice AI entirely
Limited-volume teams may find human handling or simple routing more cost-effective, but this depends on availability needs and complete costs rather than a universal call-count threshold.
High-stakes or emotionally sensitive workflows require careful risk assessment, human escalation, and appropriate professional oversight. If required consent, recording controls, privacy safeguards, or reliable escalation cannot be established, defer the relevant automated workflow.
After-hours answering or booking can provide a bounded pilot, but neither is automatically low-risk in every setting. Test realistic failures and the task outcome before expanding use.
Final verdict: compare Retell for configurable pay-as-you-go agents, Vapi for provider flexibility, and enterprise offerings for supported deployment requirements. Raw API and orchestration fees require additional components. Choose on complete cost, validated workflow behavior, controls, and the applicable contract rather than an unsupported overall ranking.
Sources
- Retell AI - AI Phone Agent Pricing (accessed October 10, 2026) - https://www.retellai.com/pricing
- Retell AI - vCX-Hard benchmark: testing 27 models (accessed October 10, 2026) - https://www.retellai.com/blog/vcx-hard-benchmark
- Retell AI - Blog / product overview (accessed October 10, 2026) - https://www.retellai.com/blog
- Vapi - Pricing: model pass-through at cost (accessed October 10, 2026) - https://vapi.ai/pricing
- Bland AI - Pricing: bundled per-minute plans (accessed October 10, 2026) - https://www.bland.ai/pricing
- Bland AI - Documentation index / llms.txt (accessed October 10, 2026) - https://docs.bland.ai/llms.txt
- ElevenLabs - ElevenAgents pricing (accessed October 10, 2026) - https://elevenlabs.io/pricing/agents
- ElevenLabs - Pricing overview (accessed October 10, 2026) - https://elevenlabs.io/pricing
- Twilio - Conversational AI pricing (accessed October 10, 2026) - https://www.twilio.com/en-us/products/conversational-ai/pricing
- Twilio - Conversation Relay product page, latency benchmarks (accessed October 10, 2026) - https://www.twilio.com/en-us/products/conversational-ai/conversationrelay
- Twilio - Conversational AI platform overview (accessed October 10, 2026) - https://www.twilio.com/en-us/products/conversational-ai
- OpenAI - GPT-Live-1 model and session pricing - https://developers.openai.com/api/docs/models/gpt-live-1
- OpenAI - Voice agents guide (accessed October 10, 2026) - https://platform.openai.com/api/docs/guides/voice-agents
- OpenAI - Voice latency and cost optimization (accessed October 10, 2026) - https://platform.openai.com/api/docs/guides/voice-latency-cost
- Google - Gemini API pricing (Live audio) (accessed October 10, 2026) - https://ai.google.dev/gemini-api/docs/pricing
- PolyAI - The technology behind the Agentic Dialog Platform (accessed October 10, 2026) - https://www.poly.ai/technology
- PolyAI - Pricing (per-minute, custom) (accessed October 10, 2026) - https://www.poly.ai/pricing
- SoundHound AI - Newsroom (LivePerson acquisition, OASYS Edge, Q2 revenue) (accessed October 10, 2026) - https://www.soundhound.com/newsroom/
- SoundHound AI - Company overview / platforms (accessed October 10, 2026) - https://www.soundhound.com/
- SoundHound - LivePerson acquisition completion, September 4, 2026 - https://investors.soundhound.com/news-releases/news-release-details/soundhound-ai-completes-acquisition-liveperson-creating-world
- TechCrunch - “Encore AI raises $30M to build AI agents that learn from customer calls” (July 29, 2026) - https://techcrunch.com/2026/07/29/encore-ai-raises-30m-to-build-ai-agents-that-learn-from-customer-calls/
- Regal - Voice AI agent platform overview (accessed October 10, 2026) - https://www.regal.ai/
- Regal - Pricing (quote-based) (accessed October 10, 2026) - https://www.regal.ai/pricing
- Observe.AI - Agentic CX platform overview (accessed October 10, 2026) - https://www.observe.ai/
- Crescendo - AI-native CX platform (accessed October 10, 2026) - https://www.crescendo.ai/
- Crescendo - Pricing (quote-based) (accessed October 10, 2026) - https://www.crescendo.ai/pricing
- FCC - “Stop Unwanted Robocalls and Texts” (robocall and AI-voice consent rules) (accessed October 10, 2026) - https://www.fcc.gov/robocalls
- Twilio - Gartner Magic Quadrant for CPaaS 2026 (cited on Twilio site) (accessed October 10, 2026) - https://www.twilio.com/en-us/report/gartner-mq-cpaas-2026
- OpenAI - Audio and voice build guide (accessed October 10, 2026) - https://platform.openai.com/api/docs/guides/audio
- OpenAI - Realtime API getting started (accessed October 10, 2026) - https://platform.openai.com/api/docs/guides/realtime
- Twilio - Voice AI use case overview (accessed October 10, 2026) - https://www.twilio.com/en-us/use-cases/voice-AI
- PolyAI - Integrations (Salesforce, Zendesk, NICE, Five9) (accessed October 10, 2026) - https://www.poly.ai/integrations
- PolyAI - Sales and service use cases for the agentic platform (accessed October 10, 2026) - https://www.poly.ai/use-cases
- Observe.AI - AI agents for customers (accessed October 10, 2026) - https://www.observe.ai/ai-agents/customers
- PolyAI - Trust centre: certifications, data residency and retention controls (accessed October 10, 2026) - https://www.poly.ai/trust
- Google - Gemini Live API overview (accessed October 10, 2026) - https://ai.google.dev/gemini-api/docs/live
- OpenAI - Realtime WebRTC transport guide (accessed October 10, 2026) - https://platform.openai.com/api/docs/guides/realtime-webrtc
- Vapi - Voice AI agents platform overview (accessed October 10, 2026) - https://vapi.ai/
- ElevenLabs - ElevenAgents product overview (accessed October 10, 2026) - https://elevenlabs.io/agents
- Twilio - Voice Conversational AI product family overview (accessed October 10, 2026) - https://www.twilio.com/en-us/voice/conversational-ai
- PolyAI - AWS SageMaker availability announcement for Dialog-RSN-1 (October 6, 2026) - https://www.poly.ai/blog/dialog-rsn-1-aws-sagemaker
- Regal - AI voice agent compliance guidance (accessed October 10, 2026) - https://www.regal.ai/blog/ai-voice-agent-compliance
- Retell - Introducing Retell Workflows, September 1, 2026 - https://www.retellai.com/blog/retell-workflows
- PolyAI - Dialog-RSN-1 vendor benchmark - https://www.poly.ai/dialog-rsn-1
- FCC - AI voice declaratory ruling, FCC 24-17 - https://docs.fcc.gov/public/attachments/FCC-24-17A1_Rcd.pdf
