The best AI voice agent platforms sort by who runs them. Dograh, the open-source AI voice agent platform we build, comes first for teams that want to self-host or go on-prem, and you can also run it on Dograh Cloud. Vapi, Retell AI and Bland AI are hosted developer platforms, and ElevenLabs Agents, PolyAI and Synthflow are managed services.
Key Takeaways
- Sort platforms by who runs them before comparing any minute rate.
- The same 10,000 minutes cost $100 to $1,999 in platform or bundled fees.
- Frameworks and speech APIs look low because your team builds the rest.
A list that mixes platforms, frameworks and speech APIs compares things that do different jobs. We sort the platforms by who runs them, from open-source Dograh to fully managed services, price one month of calls on each, then show what frameworks and speech-only vendors leave for you to build.
How we ranked the platforms
We ranked on four things a buyer can check. The first is the license, whether you can read and change the code. The second is where the platform can run, on your servers, in your cloud account or only in the vendor's cloud. The third is who runs the agent day to day, your developers or the vendor's team. The fourth is the platform fee for the same 10,000 minutes.
We left out claims a buyer cannot check, such as each vendor's own latency figures. Speech-to-text and voice vendors such as Deepgram and Cartesia, and frameworks such as Pipecat and LiveKit Agents, get their own section near the end, because each one covers only part of what a platform does.
AI voice agent platforms by who runs them
The first question is who runs the platform, because it decides where your call audio goes and who can change the agent. The table puts every option in one place before the sections cover each group.

| Platform | Who runs it | Code | Where it can run | How it bills |
|---|---|---|---|---|
| Dograh | You, or Dograh on Dograh Cloud | Open source, BSD 2-Clause | Your servers, Dograh Cloud, on-prem on Enterprise | No per-minute platform fee self-hosted, 1¢/min platform fee on Cloud |
| Vapi | Your developers on Vapi's cloud | Closed | Vapi's cloud | $0.05 a minute hosting, models and phone apart |
| Retell AI | Your team on Retell's cloud | Closed | Retell's cloud, with VPC or on-prem through sales | $0.055 a minute voice infrastructure, models apart |
| Bland AI | Your team on Bland's cloud | Closed | Bland's cloud, on-prem or VPC on Enterprise | $0.14 or $0.12 a minute with its own models |
| ElevenLabs Agents | Your team on ElevenLabs' cloud | Closed | ElevenLabs' cloud | Plans with minutes, then $0.08, language model apart |
| PolyAI | Your team or PolyAI's | Closed | PolyAI's cloud | Plans from $49 a month, or 9.5c a minute and up |
| Synthflow | Your team with Synthflow's onboarding | Closed | Synthflow's cloud | Annual enterprise agreement |
| LiveKit Cloud | Your developers build the product | LiveKit Agents framework, Apache-2.0 | LiveKit Cloud or self-hosted | Agent minutes, from $0 for 1,000 |
| Pipecat Cloud | Your developers build the product | Pipecat framework, BSD-2-Clause | Pipecat Cloud or self-hosted | $0.01 per active agent minute |
Dograh, open source you can run anywhere
Dograh ranks first on these criteria because it is the only platform in the list whose whole product is open source, and the only one you can run on your own servers without a sales contract. Self-hosting runs on Docker with at least 8 GB of memory and 4 vCPUs. Dograh Cloud runs it for you, and Enterprise adds on-prem or private-cloud deployment.
Under the call, you choose the models. The platform connects to providers such as Deepgram, ElevenLabs, Cartesia and AssemblyAI on your own keys through Model Configurations, runs open models on your own hardware, or uses speech-to-speech models such as OpenAI Realtime and Gemini Live. Dograh is the orchestration layer that runs the call flow, tools and telephony, and the speech and language models are a separate layer you pick.
New agents start in the Agent Builder. You describe the agent on the "Create an agent" page, answer any questions, approve the "Your agent requirements" card, and the agent is built for review, with "Test agent" publishing version 1. On Cloud that use is billed in Dograh credits, and a self-hosted install needs a Dograh Service Key for it, or your team builds on the canvas.
Calls run on Twilio, Vonage, Plivo, Telnyx, Exotel, Vobiz, Cloudonix or Asterisk ARI. Transfer Call works on Twilio, Telnyx, Plivo, Cloudonix and Asterisk ARI, and Dograh calls the destination first, so the caller is connected only once someone answers. A self-hosted install runs 10 calls at once by default, raised with one environment setting, while Cloud's pay as you go includes 10, with more on request.
Each finished call keeps a trace of every turn, and the QA step scores each step of test and live calls on a sample rate you set.
Self-hosted Dograh has no per-minute platform fee, so your spend goes on servers, model keys and phone minutes, which still cost money. For the agent's most common lines, pre-recorded clips play a real human recording and trim the voice bill, with text-to-speech covering the rest. Dograh Cloud's pay as you go adds a 1¢-a-minute platform fee to the model and carrier bills.
Running it yourself is a real job. The server, the database and the upgrades belong to your team, help on the open-source edition comes from the community in Dograh's Slack, and the Enterprise plan brings custom SLAs with a dedicated account manager.
Open Source Alternative to Vapi / Retell
Self-hosted voice agent platform — no per-minute fees
dograh-hq/dograh
Star on GitHub
Hosted developer platforms: Vapi, Retell, Bland
Hosted developer platforms run the calls in their own cloud while your developers set up the agent, and each prices the minute differently. Vapi charges $0.05 a minute for hosting, with models passed through at cost and the phone billed by your provider. Without a plan you get 4 calls at once, and Core at $29 a month raises that to 10. HIPAA support is a $2,000-a-month add-on, and Vapi does not support on-premise deployment.
Retell AI charges $0.055 a minute for its voice infrastructure, and complete agents run $0.07 to $0.31 a minute, depending on the models. Every account includes 20 calls at once, with more at $8 each a month, and Retell's own numbers add $0.015 a minute, with no charge on your own SIP trunk. Enterprise adds a dedicated server as an add-on, priced through sales, and Retell's security post says enterprises can also choose VPC or on-prem deployment.
Bland AI bundles its own models into the minute. Bland's Start plan costs $0.14 a minute with 10 calls at once and 100 calls a day, and Build costs $299 a month plus $0.12 a minute with 50 at once and 2,000 a day. Enterprise lifts the daily call cap and offers on-prem or VPC, and telephony is billed separately on your carrier or Bland's.
Managed platforms: ElevenLabs, PolyAI, Synthflow
Managed platforms suit teams that want more of the stack decided for them. ElevenLabs Agents sells plans with minutes included, such as Pro at $99 for 1,238 minutes, Scale at $299 for 3,738 and Business at $990 for 12,375. Each extra minute costs $0.08, and the language model is billed separately. ElevenLabs offers no self-hosted option for it.
PolyAI now sells self-serve plans next to its enterprise work. PolyAI's plans start with Essentials at $49 a month for 250 minutes, then Plus at $199 for 1,000 and Scale at $299 per location for 1,500. A Platform plan starts from 9.5 cents a minute, covering the platform, its own Dialog-RSN-1 model and the voice, and you bring your carrier over SIP at no charge.
Synthflow sits at the other end. It is sold as an annual enterprise agreement with no rate by the minute and no free plan or trial, so the price comes out of a scoping call with its team.
What 10,000 minutes cost on each
A worked example shows how far apart the bills sit. Take 10,000 call minutes in one month and price each platform's own line, then list what you still pay for on top.
| Platform | Platform or bundled line for 10,000 minutes | Still to add |
|---|---|---|
| Dograh, self-hosted | No per-minute platform fee | Servers, model keys or local models, phone minutes |
| Dograh Cloud | 10,000 x $0.01 = $100 | Models, on your keys or Dograh's, and the carrier |
| Vapi | 10,000 x $0.05 = $500 | Models, carrier, $29 Core above 4 calls at once |
| Retell AI | 10,000 x $0.055 = $550 | Models and voice ($700 to $3,100 all-in), $150 on Retell numbers |
| Bland AI Start | 10,000 x $0.14 = $1,400 | Phone minutes. Fits only if calls average over about 3.3 minutes |
| Bland AI Build | $299 + 10,000 x $0.12 = $1,499 | Phone minutes |
| ElevenLabs Agents Scale | $299 + 6,262 x $0.08 = $799.96 | Language model and carrier |
| ElevenLabs Agents Business | $990, with 12,375 minutes included | Language model and carrier |
| PolyAI Platform | 10,000 x $0.095 = $950 and up | Your carrier |
| PolyAI Scale, one location | $299 + 8,500 x $0.20 = $1,999 | Your carrier |
| Synthflow | Priced in a scoping call | Set in the agreement |
The rows are not like for like. Bland's and PolyAI Platform's lines already include their models, while Vapi, Retell's $550 line and ElevenLabs still need them, and a large reasoning model can add more than the platform line itself. Bland Start also hides a limit, since 100 calls a day is about 3,000 calls a month.
Price the model you plan to run before you compare. Plan caps, add-ons and the cost of moving away add their own lines, which our post on the hidden costs of closed voice AI platforms works through.
Join the Dograh Community
Dograh is an OSS alternative to Vapi. Join our Slack community for queries, releases, best practices & community interactions.
Frameworks and speech APIs are not platforms
An agent framework gives your developers the code to run a voice pipeline, and the product around it is theirs to build. LiveKit Cloud hosts the Apache-2.0 LiveKit Agents framework, with Build free for 1,000 agent minutes and Ship at $50 a month for 5,000, then $0.01 a minute. That puts 10,000 minutes at $100. Pipecat Cloud hosts the BSD-2-Clause Pipecat framework at $0.01 per active minute on its smallest agent size, also $100, with Daily's phone minutes at $0.018.
Those numbers look low because the product is not built yet. The builder, the phone setup, transfers, campaigns, QA and call history are still on your team's list. Our guide to the best open-source voice AI platforms lists what a team builds on a framework before its first real phone call.
Speech vendors cover one layer only. Deepgram and Cartesia sell speech-to-text or voices, and ElevenLabs sells voices apart from its agent product, so each one needs a platform or framework around it to hold a call. Our list of Deepgram alternatives compares speech APIs with full agent platforms.
Which platform fits your team
If open code, self-hosting or on-prem matters, Dograh is the first platform to try, on your servers or on Dograh Cloud while you decide. A developer team that wants a vendor to run the calls will look at Vapi or Retell AI, and a phone team that wants one bundled rate and no model choices will look at Bland AI.
ElevenLabs Agents suits a brand whose voice is part of the product. PolyAI and Synthflow suit teams that want the vendor's people involved in the build. Frameworks suit a company that wants to build its own voice product and has the engineers for it. If you expect large spikes, compare how each one handles calls at once, which our post on voice agent platforms for high call volume covers platform by platform.
Whichever you shortlist, run the pilot over a real phone number. A browser demo runs at 16 kHz, while most phone calls still arrive as 8 kHz audio, so an agent that sounds right in the dashboard can mishear names and numbers once it answers the phone.
Glossary
- Platform fee
- The per-minute charge for running the call on a platform, separate from the speech and language models and the phone line. Vapi's is $0.05 a minute and Dograh Cloud's is 1 cent.
- Daily call cap
- A limit on how many calls a plan can start in one day, separate from how many run at once. Bland's Start plan allows 100 a day and Build 2,000.
- Agent framework
- Code your developers build a voice agent with, such as Pipecat or LiveKit Agents. It runs the pipeline, while the builder, phone setup and QA are left to your team.
- Orchestration layer
- The part of a voice agent that runs the call flow, tools and telephony and passes audio between models. Dograh is this layer, and the models under it are a separate choice.

