Vapi Review 2026

Vapi is worth a shortlist when voice is product infrastructure, not a side experiment. The official platform is API-first: configure assistants, wire STT, LLM, and TTS providers, connect telephony through Vapi or your carrier, attach server-side tools, and monitor calls in a unified dashboard. Vapi markets sub-500ms average latency, 99.9% uptime for enterprise clients, and deployments from startups to operators handling millions of calls per month. Customer stories on the homepage include Ring, Kavak, and healthcare operators at scale.

The buying mistake is treating Vapi hosting at $0.05 per minute as the finished call cost. On the Build tier, STT, TTS, LLM, and telephony transport are billed at provider cost on top of Vapi hosting. Bring-your-own API keys can zero out some Vapi-side model charges, but carrier minutes and transport still apply. SMS and chat hosting is $0.005 per message. Concurrency is capped at ten included lines on Build, with extra lines at $10 per line per month. Scale moves to annual contracts with committed volume, SOC 2, HIPAA, PCI, SSO, RBAC, data residency, and dedicated support.

How we evaluated this

This is a desk review, not a paid latency benchmark across carriers. We checked the official homepage, pricing matrix, GDPR documentation, and the usage calculator on August 20, 2026. We cross-read the AI phone agents buyer guide on this site, which already modeled Vapi against Retell, Bland, and receptionist vendors. We did not independently audit SOC 2 Type II, complete a HIPAA BAA review, run a multi-week outbound dialer test, or measure word-error rates on your accent mix.

We weight API control and provider flexibility, true all-in minute economics, concurrency and peak-call behavior, transfer and barge-in reliability, tool-call success on live audio, retention defaults on Build, and compliance packaging on Scale. Vapi markets AI guardrails, enterprise SSO, and SOC 2, HIPAA, and PCI compliance on Scale. Treat HIPAA at $2,000 per month and Zero Data Retention at $1,000 per month as contract checkpoints on Build, not implied inclusions.

Cost model

The $0.05 hosting line is one layer in a four-provider stack

Vapi looks inexpensive because Build advertises $0.05 per minute for platform hosting. The units that actually constrain a production rollout are STT minutes, LLM tokens during the realtime loop, TTS characters or seconds, telephony transport from Twilio or your BYOC carrier, and concurrency lines during peak inbound or outbound bursts. That is a different shape from Goodcall at $79 per agent with unlimited minutes, or Bland at a bundled connected-minute rate where STT, TTS, and LLM sit inside one vendor number.

• Hosting is not the call. Vapi charges $0.05 per minute for container hosting on voice calls. Model provider costs for STT, LLM, and TTS are at cost, or $0 to Vapi when you bring your own API key. Transport is charged by the telephony provider, not Vapi. Budget four line items, not one.
• SMS and chat are separate meters. Build lists $0.005 per message for SMS or chat hosting, with model costs also at provider rates. A voice-first team that adds post-call SMS confirmations needs both meters in the spreadsheet.
• Concurrency is a capacity gate. Build includes ten concurrent call lines. Line eleven costs $10 per month per extra line. A marketing burst that drives fifty simultaneous inbound calls is a capacity purchase, not a surprise you discover on launch day.
• Retention is short on Build. Call history is fourteen days and chat history is thirty days on the public pricing matrix. Long-running QA, dispute review, or compliance archives may require Scale, exports, or your own warehouse.
• Compliance is priced explicitly. HIPAA add-on is $2,000 per month on Build and Scale. Zero Data Retention is $1,000 per month. Neither is bundled into the $0.05 hosting rate.

Worked example on public USD rates checked August 20, 2026. A SaaS support line averages four minutes per call and handles 5,000 connected minutes per month. Vapi hosting at $0.05 is $250. Assume $0.01 per minute for STT, $0.03 for LLM inference across the turn loop, $0.04 for TTS, and $0.008 for carrier transport. Stack total near $0.128 per minute, or $640 for the month before engineering time. Add two extra concurrency lines at $10 each if peaks require twelve simultaneous calls: $20 more. If the same program needs HIPAA, add $2,000 per month for the add-on. The pilot that looked like $250 on the pricing page is closer to $640 to $2,860 depending on stack choices and compliance scope.

Contrast with a bundled competitor from the phone agents guide. Bland Build at $299 per month plus $0.12 per connected minute on 5,000 minutes is $899 before transfers. Vapi can be cheaper or more expensive depending on provider choices and eng ownership. The decision is control versus bundled simplicity, not a single public rate card comparison.

Who Vapi fits best

Surfaces to verify

Vapi surfaces are telephony, realtime voice loops, SMS or chat, APIs, and downstream integrations. They are not a WhatsApp support inbox or Instagram DM stack. Map each path before you promise omnichannel coverage to stakeholders.

• PSTN inbound and outbound: confirm number provisioning, caller ID, recording announcements, and jurisdiction-appropriate consent language on live calls.
• SIP and BYOC: validate trunk authentication, codec behavior, and failover if your carrier drops mid-call.
• Web and mobile SDK voice: test browser mic permissions, mobile network handoffs, and time-to-first-word on 4G and 5G, not only office Wi-Fi.
• SMS and chat channel: Build lists $0.005 per message hosting. Confirm model costs, opt-out handling, and whether chat history at thirty days meets your QA needs.
• Server-side tools and webhooks: calendar writes, CRM lookups, ticket creation, and custom APIs during the call. Test timeout behavior when your backend is slow.
• Warm and cold transfer: human handoff to queues, mobiles, or third-party contact centers. Demos often skip angry-caller transfer stress.
• Assistant dashboard and API: version prompts, swap models, and run test calls before promoting config to production traffic.
• Observability exports: transcripts, recordings, latency metrics, and call logs. Confirm fourteen-day retention on Build and export paths to your warehouse.
• Downstream orchestration: pair with YourGPT AI or your CRM when post-call summarization, approval gates, or ticket schemas need governed handoff beyond raw webhooks.

Voice workflow and handoff

The intended loop is: a call arrives via PSTN or SIP, Vapi runs the STT to LLM to TTS realtime loop with barge-in, the assistant calls tools for calendar, CRM, or status lookups, and the call ends with a structured outcome, transfer, or follow-up SMS. The buying question is whether that loop survives peak concurrency, transfer edge cases, backend tool latency, and your compliance rules.

Inbound support teams often start with after-hours capture, appointment reschedule, or FAQ with hard knowledge bounds. Outbound teams may run consented reminders or no-show recovery. Each path needs different suppression, consent storage, and transfer behavior. Vapi provides the runtime. Your team owns the policy graph.

Handoff quality separates a demo from operations. Test warm transfer briefs, cold transfer to a busy queue, caller hang-up during tool execution, and recovery when the LLM chooses the wrong tool. Test post-call objects: does your webhook receive structured fields a human can trust, or a prose summary someone must rewrite before CRM insert? If humans pick up, measure repeat-myself complaints and abandoned-after-transfer rate.

Pair voice runtime with a control layer when policies span channels. YourGPT AI fits when the same business rules must govern chat, email, and phone, and when approval gates block unsafe refunds or account changes. Vapi is the telephony and realtime loop. It does not replace governed knowledge and escalation design.

Pricing checked August 20, 2026

Vapi pricing: Build usage meters and Scale enterprise contracts

Public USD packaging is Build (usage-based, self-serve signup) and Scale (annual contract, contact sales). Rates below are from vapi.ai/pricing. Model provider and telephony transport costs are additional on Build unless you bring keys and carriers that zero out specific lines.

Source: vapi.ai/pricing, checked August 20, 2026. Confirm checkout credits, included minute grants, provider rates, carrier transport, concurrency needs, and add-on scope before budgeting. Scale pricing requires a sales quote.

AI capability: realtime loop, tools, and guardrails

Vapi orchestrates the voice AI stack rather than locking you to one model vendor. You configure STT for accurate barge-in, an LLM for policy and tool selection, and TTS for natural speech. The platform advertises AI guardrails to reduce hallucinations and protect data integrity during live conversations. On Scale, priority model and provider access can matter when you need approved subprocessors or private endpoints.

Tool calling during audio is the product value. Assistants can hit your APIs for calendar availability, order status, ticket creation, or internal lookups while the caller stays on the line. Latency compounds: STT finish, LLM plan, tool round trip, TTS start. Sub-500ms marketing targets assume tight provider choices and healthy backend responses. Test with your slowest integration, not a mock endpoint.

Provider flexibility helps security reviews. If your InfoSec team requires Azure OpenAI, a specific Deepgram region, or ElevenLabs with a DPA, Vapi supports wiring those choices explicitly. That same flexibility means someone must own model routing, fallback when a vendor errors, and cost caps when GPT-4 class models run on high-volume inbound traffic.

Feature areas to verify in a pilot

• Mobile PSTN call with p50 and p95 time-to-first-word on your chosen STT, LLM, and TTS stack.
• Barge-in and interruption: caller talks over the agent mid-sentence without loop failure.
• Warm transfer to a human with a spoken brief; cold transfer to a queue that may not answer.
• Tool call that writes to calendar or CRM, including failure when the API times out.
• Outbound consented call with opt-out captured and suppression verified within minutes.
• Peak concurrency at eleven simultaneous calls to validate line overage billing and audio quality.
• SIP or BYOC path if enterprise telephony is required, including failover behavior.
• Recording and transcript export before fourteen-day retention expires on Build.
• Provider outage drill: swap STT or LLM config and confirm recovery without redeploying telephony.
• Post-call webhook schema validated against your ticket or CRM required fields.

Analytics and operating visibility

Vapi markets real-time monitoring and continuous improvement loops across calls. Build includes dashboard visibility with fourteen-day call history and thirty-day chat history per the pricing matrix. That is enough for active sprint QA, not always enough for quarterly coaching, dispute review, or compliance audits without exports.

Operationally, track four numbers during pilot: call completion rate, transfer success rate, tool-call success rate, and median end-to-end latency from connect to first agent speech. Add cost per successful outcome once provider invoices arrive. Without those, you cannot compare Vapi stacks to bundled vendors fairly.

Scale adds enterprise-grade uptime SLA, custom support SLA, and dedicated account team coverage. Ask for sample operational reports during sales review if managers need talk-time, intent, or failure-code visibility across sites. Export to your warehouse early if BI owns reporting.

Security, data handling, and compliance

Vapi publishes GDPR documentation at docs.vapi.ai/security-and-privacy/GDPR, describing encryption in transit and at rest, access controls, subprocessors such as Stripe and analytics providers, and data subject rights. Scale advertises SOC 2, HIPAA, PCI, SSO, OAuth, RBAC, and data residency. Build does not list SOC 2, SSO, or RBAC on the public pricing matrix.

HIPAA at $2,000 per month and Zero Data Retention at $1,000 per month are explicit add-ons on both tiers. Obtain BAA scope, subprocessors for STT and TTS vendors you select, and retention behavior with and without the ZDR add-on before connecting PHI or financial data. Default fourteen-day call history on Build is not a long-term archive strategy.

Reconcile marketing copy with your order form. Ask who can access recordings, whether provider logs retain audio, how deletion propagates to webhooks you stored, and whether PCI scope applies if payments are discussed on calls. Pair with counsel for outbound TCPA and artificial-voice rules before scaling dialers.

What questions should you ask before buying Vapi?

• What is the all-in per-minute cost at our expected STT, LLM, TTS, and carrier choices on 5,000 and 50,000 minutes?
• How many concurrent lines do peak inbound and outbound campaigns require, and what is the overage line cost?
• Which subprocessors apply when we bring our own keys versus Vapi-billed provider paths?
• Does Build retention meet QA and compliance needs, or do we need Scale or Zero Data Retention?
• What is included in the HIPAA add-on, and will you sign a BAA before PHI testing?
• How do warm and cold transfers behave when the human queue is full or the callee hangs up mid-brief?
• What happens to tool calls when our backend exceeds a latency threshold during live audio?
• Can we export transcripts and recordings automatically before fourteen-day deletion on Build?
• What Scale committed-volume pricing applies at our twelve-month minute forecast?
• Who owns weekly transcript review, prompt change control, and incident response on our side?

What red flags should you watch for with Vapi?

• Finance approved budget using $0.05 per minute without STT, LLM, TTS, or carrier line items.
• The demo ran only on web SDK over Wi-Fi, not PSTN mobile calls your customers use.
• Transfer success was never tested with an upset caller or a busy human queue.
• Outbound dialing starts before consent storage and suppression latency are proven.
• Healthcare or PCI scope is assumed from marketing icons without HIPAA or ZDR add-ons in the quote.
• Leadership expects a managed receptionist outcome without staffing engineers for QA and prompt versions.
• Post-call CRM writes are prose summaries with no schema validation or approval gates.

What are the best alternatives to Vapi?

Pick by whether you need programmable voice infrastructure or a packaged outcome. Retell AI, Goodcall, and Smith.ai are covered in the AI phone agents buyer guide. Our Synthflow AI review and Bland AI review cover enterprise contracts and bundled phone-agent minutes. We link live reviews below.

• YourGPT AIChoose YourGPT when omnichannel support, governed knowledge, approval gates, and post-call orchestration matter alongside voice, and you will pair it with a telephony runtime.
• Intercom FinChoose Intercom Fin when your team already runs Intercom and wants AI across chat, email, and voice inside one inbox and helpdesk data model.
• Fireflies.aiChoose Fireflies when the job is meeting capture and conversation intelligence, not PSTN phone agents with tool calling during live calls.
• TidioChoose Tidio when you need affordable web chat and light automation, not SIP depth or custom voice stacks.
• Synthflow AIChoose Synthflow when you need enterprise telephony, BPO-scale deployment, and annual contract scope rather than self-serve API minutes.
• Bland AIChoose Bland when you want a phone-first agent builder with pathway testing, bundled AI minutes, and optional platform tiers without assembling STT, LLM, and TTS yourself.

For Retell, Bland, Goodcall, and Smith.ai comparisons, use the AI phone agents buyer guide and model bundled per-minute or per-agent pricing against your stack math.

Workflow test

What Vapi needs to prove in a real workflow

A clean web-sdk demo is not evidence. Run this four-step test on your numbers and policies.

1. Call on the real network.Place ten PSTN calls from mobile handsets on cellular data. Measure time-to-first-word, barge-in, and audio drop rate.
2. Execute the tool path.Run calendar write, CRM lookup, and ticket create flows with your slowest backend. Confirm timeout speech and retry behavior.
3. Transfer under stress.Warm transfer to a human who is busy, cold transfer to a queue, and caller hang-up mid-transfer. Score repeat-myself complaints.
4. Reconcile the bill.After 500 connected minutes, compare Vapi hosting, provider, transport, and concurrency invoices to the pre-pilot model.

• AI phone agents buyer guideCategory context for inbound versus outbound, Retell and Bland pricing shapes, and transfer pressure tests.
• YourGPT AI reviewCompare omnichannel control and workflow execution when voice sits beside chat and email.
• Intercom Fin reviewCompare helpdesk-native AI when voice is one channel inside Intercom.
• Fireflies.ai reviewCompare meeting voice agents and conversation intelligence versus PSTN runtime infrastructure.
• Customer support AI agents guideWhere phone agents fit in support automation and handoff design.
• How to choose an AI agent platformApply meter, data-use, and workflow criteria to any agent rollout.
• Tools directoryScan adjacent voice, support, and automation tools on one list.
• AI agent buyer scorecardTurn the four-step voice workflow test into a written go or no-go.

Official product demo

See Vapi provider wiring before you trust latency claims

This official Vapi setup walkthrough shows dashboard configuration, provider connections, and test-call behavior. Treat it as a product tour, not independent proof of transfer reliability or your all-in minute cost. After watching, repeat the workflow on PSTN with your STT, LLM, TTS, and telephony stack.

Watch on YouTube and rerun the same test call on your carrier and provider stack.

Claim and source ledger

Should you choose Vapi?

Vapi is a developer-first voice AI platform for PSTN, SIP, and SDK calls with configurable STT, LLM, and TTS providers, server-side tools, and enterprise Scale packaging. Public Build pricing as of August 20, 2026 lists $0.05 per minute for call hosting and $0.005 per message for SMS or chat hosting, with provider and transport costs additional, ten concurrent lines included, and HIPAA at $2,000 per month as an add-on.

Choose Vapi when voice is product infrastructure, you need provider and telephony control, and your team will own QA, prompt versions, and stack math. Look elsewhere when you wanted a managed receptionist, when YourGPT AI or Intercom Fin already covers your channel mix, when outbound compliance is unresolved, or when $0.05 per minute was mistaken for the finished bill. Run the four-step workflow test on PSTN, not on a dashboard demo alone.