Best AI Voice Agent Platforms in 2026, Ranked by Who Owns Your Phone Numbers When You Leave

When an AI voice platform closes, gets acquired, or raises its per-minute rate past your margin, the telephony layer is the exit variable no pricing page will tell you about. The phone numbers your agents answer — the ones your clients call, the ones baked into their contact flows and their customer records — may sit in your own carrier account, or in the platform's. That single question determines whether leaving costs an afternoon of config work or a prolonged negotiation to reclaim what should have been yours.

Callia is a multi-tenant AI voice agent platform sold once as source code. You deploy it under your own brand, connect your own telephony provider account (Twilio, Telnyx, Plivo, or Vonage), and charge your own clients for AI phone agent service on plans you configure. The carrier relationship is yours at deployment and yours when you stop. This guide maps the category by that variable — and names the right answer for every buyer, including the ones where Callia is not it.

Five things to evaluate before you choose

Telephony portability and exit cost. The organizing question for this guide. A phone number held in your own Twilio or Telnyx account moves with you when you replace your platform software; the number stays, only the code changes. A number provisioned inside a platform's own carrier relationship requires that platform's cooperation to port — or you re-provision, and your clients dial a new number. None of the pricing pages in this category address this question directly. Several platforms in this guide support bringing your own carrier (documented below, per primary sources); several do not disclose their policy, and "unpublished" is what this guide calls it rather than inferring.

Per-minute economics — with the counterweight that makes the argument honest. This category bills by the minute, and the cost of a call splits into two layers: the platform's orchestration margin, and the underlying call stack — telephony, STT, TTS, and a language model. Self-hosting removes the margin. It does not remove the call. A mid-range self-hosted stack runs roughly $0.068/min on Twilio, Deepgram, Cartesia-class TTS, and a GPT-4.1-class model — from each vendor's published rates as of 2026-08-07. The saving percentage varies by configuration: roughly 73% against a lean stack, roughly 24% against a premium setup with a frontier model and ElevenLabs voice — because the irreducible provider cost dominates at that tier. The argument is strongest at high call volumes and weak below 1,000 minutes a month.

Code and data ownership. The platform is rented or owned. Call transcripts, recordings, and client data live on the vendor's infrastructure or on yours. Vapi's Build tier retains call history for 14 days; Zero Data Retention is a $1,000/month add-on. For a buyer whose business model is the platform, the data is the business.

Multi-tenant operator capability. The major hosted platforms — Vapi, Retell, Bland, and ElevenLabs Agents — are not sold as platforms you white-label under your own brand name and operate for your own clients at your own prices. If that is the goal, you are looking at a different tier of the market: white-label services with recurring platform fees, or a purchased codebase.

Deployment burden, honestly. Self-hosting is not cost-free. A persistent-WebSocket relay service for live calls, a Supabase Postgres database, Vercel deployments, and a QStash account are infrastructure costs and ongoing maintenance responsibilities. A hosted platform at $0.13/min is not more expensive than self-hosting in every configuration at every volume — the counterweight is real and it scales down with your minute count.

---

Tier 1 — Vendor-provisioned telephony

These platforms provision and manage phone numbers within their own infrastructure. What telephony portability looks like when you leave — whether porting is possible, how long it takes, and who initiates it — is not publicly documented by the vendors in this tier. Verify before building a business-critical number on their provisioning layer.

ElevenLabs Agents

ElevenLabs is a voice-model company that has extended its TTS and STT technology into a full agent layer for phone calls. The pricing model bundles minutes into plan tiers: Starter at $6/month includes 75 minutes of agent calls at an effective rate of $0.08/min; the Business tier at $990/month includes 12,375 minutes, also at $0.08/min for included usage. LLM costs and telephony are charged on top of and outside the per-minute audio rate — the plan buys the voice capability, and the model and the phone line arrive as separate line items. Source: elevenlabs.io/pricing/agents, 2026-08-07.

Honest limit. No BYO telephony path is documented in ElevenLabs Agents' primary documentation as of 2026-08-07. The telephony is billed as their service; what the portability terms look like for the numbers they provision is unpublished. If you need the company's own voice models integrated without provider configuration work, ElevenLabs Agents is a coherent choice. It is not a multi-tenant operator platform, and the LLM and telephony costs layered on top of the audio plan fee make the true cost-per-minute comparison harder to complete than Retell's transparent rate card.

Callin.io

The closest thing to a white-label incumbent in the mid-market agency tier. Secondary sources suggest pricing in the range of approximately $119/month to $499/month for white-label configurations, with client caps on all tiers below the highest, and overage rates of roughly $0.10–$0.14/min — but these figures come from comparison sites rather than from callin.io's own pricing page, and were not verified against a primary source for this guide. Source: comparison site, LOW confidence; verify directly before any commercial decision.

Honest limit. Client caps on most tiers bound your resale scale. Telephony portability policies are not publicly documented. Callin.io is a service you subscribe to, not code you own — the platform, the data, and the carrier relationships stay with the vendor.

---

Tier 2 — BYO carrier supported, code stays rented

These platforms let you connect your own Twilio, Telnyx, or SIP account. That means the phone number is in your carrier account from day one. When you switch platforms, the number does not move — only the software does. What stays behind when you leave is the orchestration code, the billing and admin infrastructure, and whatever call data the vendor retains under their own retention policy.

Vapi

The most-cited name in this category. Vapi charges a platform fee of $0.05 per minute on every call it handles, on top of whatever your providers cost. If you bring your own API keys — which Vapi supports across transcription, LLM, and all voice providers, with no platform charge applied to those provider layers — you pay $0.05/min for Vapi's infrastructure and your own provider costs directly. BYO telephony is supported via provider configuration, meaning the number can live in your own carrier account from day one. Source: vapi.ai/pricing and docs.vapi.ai/customization/provider-keys, 2026-08-07.

Honest limit. HIPAA compliance is a $2,000/month add-on. Zero Data Retention is a $1,000/month add-on on top of that. Build tier call history defaults to 14-day retention. The code, the data-retention policy, and the billing infrastructure are Vapi's to change. Vapi is a rented orchestration layer, not code you own.

Retell AI

Retell publishes the most transparent rate card in this category. Every component is line-itemed, which makes honest arithmetic possible and distinguishes it from every other platform in this tier. The base fee — "voice infrastructure" in Retell's terminology — is $0.055/min. A mid-range configuration (GPT-4.1-class LLM at $0.045/min, Cartesia-class TTS at $0.015/min, Twilio telephony at $0.015/min) runs $0.130/min all-in; a budget configuration reaches $0.088/min; a premium setup with ElevenLabs voice and a frontier model reaches $0.190/min. Source: retellai.com/pricing, 2026-08-07.

BYO telephony is documented: elastic SIP trunking, with import of numbers from Twilio, Telnyx, and Vonage. Source: docs.retellai.com/deploy/custom-telephony, 2026-08-07. The number is in your carrier account.

Honest limit. TTS is menu-select rather than BYO-key. You choose from Retell's supported voice providers (Cartesia, ElevenLabs, OpenAI, Fish, Minimax); the fees are billed by Retell against your Retell account, not drawn from your own ElevenLabs or Cartesia credentials. The platform code and your call data stay Retell's.

Bland AI

Bland's Start tier carries no plan fee: $0.14/min all-in, no monthly minimum. The Build tier adds a $299/month plan fee and reduces the per-minute rate to $0.12/min — the break-even volume where the $299 plan fee pays back against the $0.02/min rate reduction is approximately 14,950 minutes per month. Below that, the Build plan buys higher concurrency and daily capacity limits, not cheaper calls. Source: bland.ai/pricing, 2026-08-07.

BYO telephony is documented explicitly: "BYOT customers do not pay transfer fees. You handle carrier costs directly and Bland charges only the per-minute AI rate." Source: bland.ai/pricing, 2026-08-07. The carrier account is yours.

Honest limit. The model layer is Bland's own — there is no path to connect a third-party LLM via your own API key. Daily caps apply on lower tiers, though whether the cap counts calls or minutes is not stated on Bland's own pricing page. The orchestration code stays Bland's.

Synthflow

Synthflow was the white-label incumbent for mid-market agencies. As of 2026-08-07, its pricing page serves an enterprise-only engagement: $30,000/year minimum, custom scope, "Contact Sales." The pay-as-you-go documentation page 301-redirects to that same enterprise page. A BYO Twilio configuration is widely reported in secondary sources with the platform fee reportedly dropping significantly on that configuration — but those component rates are no longer verifiable against a Synthflow-owned primary page and are LOW confidence. Source: synthflow.ai/pricing, fetched 2026-08-07.

Honest limit. No public self-serve pricing. If you are not a $30,000/year buyer, the pricing conversation starts at "contact sales." The white-label reseller positioning that once competed with Callin.io and direct alternatives has effectively exited the accessible tier.

---

Tier 3 — Own the stack: carrier account, code, and data

These options give you the phone number, the platform software, and the call data. Leaving means changing software on your own infrastructure — exit cost is an afternoon.

Dograh

BSD-2-Clause, free, 5,163 GitHub stars, last pushed 2026-08-07. Dograh's own repository describes itself as a "Self-hosted alternative to Vapi and Retell." It ships a visual workflow builder, BYO keys across STT, TTS, and LLM (or speech-to-speech), MCP server support, and a docker compose up deploy path. You configure your own telephony provider; the carrier account is yours. Source: github.com/dograh-hq/dograh, via GitHub API, 2026-08-07.

Honest limit. Dograh is a community project, not a finished commercial platform. There is no built-in billing engine, no DB-driven plan management, no multi-tenancy for managing multiple clients' agents in isolated workspaces, and no outbound campaign dial engine. An engineer who wants a self-hosted voice assistant for a single tenant or who intends to build the business layer themselves will find Dograh a serious starting point. A buyer who wants to deploy a named voice-AI service — with per-client plan pricing, Stripe payouts, a campaign engine, and an admin panel already wired — is looking at the build scope the product does not include.

Pipecat and LiveKit Agents

Pipecat (BSD-2-Clause, 13,978 stars, pushed 2026-08-07) and LiveKit Agents (Apache-2.0, 12,720 stars, pushed 2026-08-07) are the active open-source frameworks in this category. Pipecat assembles a frame-processor pipeline in Python — STT, LLM, and TTS in sequence, with turn detection and barge-in handling. LiveKit Agents layers agent logic on LiveKit's WebRTC server with built-in interruption handling and telephony integration. Both are maintained by commercial entities with commercial interests in the underlying infrastructure, and both are actively shipped. Source: GitHub API, queried 2026-08-07.

Your telephony account — Twilio, Telnyx, or anything you configure — is yours from the first call.

Honest limit. Both are frameworks, not products. They supply the voice pipeline. Authentication, RBAC, multi-tenancy, billing and plan management, an admin panel, contact management, and a campaign dial engine are not included — they are the work a buyer does on top before they have a product. For an engineer building something custom on a permissive license, Pipecat and LiveKit Agents are the serious options. For a buyer who wants a finished SaaS to operate, they are the materials, not the building.

Callia

Callia is $349 once — a one-time purchase, no per-minute platform fee from Vapi or Retell, no recurring subscription to the platform software. You connect your own Twilio, Telnyx, Plivo, or Vonage account; the carrier relationship is yours at deploy and yours when you stop. You deploy under your own brand; the subscription your clients pay goes into your Stripe account at prices you configure in your admin panel.

The platform ships as a complete multi-tenant SaaS: agent builder with a visual call flow designer, knowledge base with RAG retrieval (pgvector cosine similarity), outbound campaign engine backed by QStash with a DNC check before every dial, five-layer swappable provider stack (telephony / STT / TTS / LLM / middleware), multi-tenant billing through Stripe, DB-driven plan management, call logs with full transcripts and timestamp-synced audio, phone number management, contact management with CSV import via Data Bridge, an MCP server, and a Connect API.

Two operating modes. Middleware Mode routes calls through Vapi or LiveKit Agents — it runs on Vercel serverless without any extra infrastructure and you pay the middleware platform fee while in that configuration. It is the documented default and the recommended starting point. Direct Mode routes calls through your own provider stack at a lower per-minute rate; the counterweight is a separate persistent-WebSocket relay service deployed on Fly.io or Railway (approximately $5–15/month for a solo worker), because a live-call WebSocket cannot survive on Vercel's serverless function timeout. Direct Mode savings are real — in the range of $0.03–$0.07/min at mid-range configurations — but they require two running services, not one.

Honest limit. Voice only. No SMS, WhatsApp, or chat — a scope that Vapi ($0.005/message), Retell ($0.002+/message), and ElevenLabs ($0.003/message) do not share; a buyer who needs one platform for voice and messaging should not purchase Callia. Middleware Mode continues to pay a middleware platform fee. Direct Mode requires an always-on relay worker that the standard Vercel deploy does not include. There is no production track record or verifiable customer base — the major hosted platforms have public customers, uptime histories, and volume references that Callia does not have, and that is a real difference when evaluating the risk of a v1.0 build for a revenue-critical use case. The stack runs on Next.js + Supabase + Vercel + QStash and requires a direct OpenAI API key for embeddings — it is not a single binary and is not trivially portable outside that constellation.

Try the demo at https://callia.saascode.ai or purchase at https://saascode.ai/products/callia.

---

Dead entries — still appearing in 2026 comparison articles

Three platforms recur in current "best AI voice agent" listicles. A buyer who found any of them elsewhere deserves their current status.

Air.ai settled with the FTC in March 2026 — an $18M consent judgment largely suspended, founders and five related entities permanently banned from marketing business opportunities. The product was dark before the settlement was signed. Source: FTC March 2026 press release, cited by multiple outlets; ftc.gov returned 403 to automated retrieval — verify against the FTC's own case page before citing in a legal context. For buyers looking for a replacement: Vapi, Retell, and Bland are the managed-platform path; Dograh and Callia are the own-the-code path.

Vogent was acquired by Aircall on May 6, 2026. The San Francisco team and technology folded into Aircall's AI Voice Agent product. Vogent does not exist as a standalone platform. Source: Aircall newsroom, 2026-05-06.

Vocode (vocodedev/vocode-core) has not received a commit since November 15, 2024. Its last release, tagged 2024-08-07, carries a beta label. The repository is not archived, and its readme carries an open call for community maintainers — the maintainers have not declared the project dead, but no active development is taking place. Source: GitHub API, queried 2026-08-07. A dormant framework is not a platform to build a client-facing service on.

---

How to choose

If your situation is…	The right call
You want a hosted platform, own your carrier account already, and want the most transparent per-minute rate card	Retell AI — only vendor in this category publishing every component line by line
You want BYO telephony on a hosted platform with no plan fee required at low volume	Bland AI — BYOT documented, Start tier has no monthly minimum
You want BYO keys across every provider layer on a well-funded managed platform	Vapi — BYO across STT/TTS/LLM/telephony; $0.05/min platform fee is the delta
You need bundled minutes and the platform's own voice models as an integrated system	ElevenLabs Agents — but verify telephony portability directly before provisioning production numbers
You need white-label hosting for a $30,000+ annual engagement	Synthflow — but confirm current terms; no self-serve pricing as of 2026-08-07
You want a self-hosted voice pipeline on a permissive open licence, willing to build the platform layer	Dograh (BSD-2) — if you want a head-start product; Pipecat (BSD-2) or LiveKit Agents (Apache-2.0) — if you want a framework
You want to deploy and operate a named multi-tenant voice-AI service, billing your own clients, code in your hands	Callia — accepting voice-only scope, the relay-worker requirement for Direct Mode, and the absence of a production track record
You were using Air.ai, Vogent, or Vocode	All three are no longer available as standalone products (see dead entries above)

---

Honest verdict

The telephony-portability axis reorders the list in a way a per-minute rate table cannot: cheap is not the same as portable, and portable is not the same as owned. None of the pricing pages in this category says which it is.

The most durable decision in this category is not which platform has the cheapest per-minute rate. It is whether you want to operate someone else's orchestration software or your own. The hosted cohort — Vapi, Retell, Bland, ElevenLabs Agents — is the right answer for most operators who want to move fast, start without infrastructure overhead, and pay per minute for a managed service. The self-hosted cohort — Dograh for engineers building on BSD-2, Callia for operators who want a finished commercial platform with billing and multi-tenant client management built in — is the right answer for the buyer who is building a business that depends on the platform's continuity and wants the carrier relationship and the code to outlast any vendor decision.

Callia earns its position in this guide for the specific buyer who wants to run a named, multi-tenant voice-AI service — their own brand, their own prices, their own Stripe account — on a codebase and a carrier relationship that belong to them. For every other reader in this table, the honest verdict routes somewhere else, and each section has said so.

---