ElevenLabs Agents vs Retell AI: The Real Per-Minute Cost

ElevenAgents plans are billed on call minutes, separately from the credit pool the rest of the ElevenLabs product uses. Each tier bundles minutes and a concurrency ceiling:

Go past your included minutes and you pay $0.08 a minute. Go past your concurrency and you pay $0.16. Text messages are $0.003 each.

None of that includes the model or the phone line.ElevenLabs is explicit about it: the LLM and any telephony are billed separately, on top, at the provider's rates. So "$0.08 a minute" is the hosting fee for the agent, not the cost of running it. Add a mid-tier model at roughly $0.02 a minute and a US phone number at roughly $0.014, and the honest figure is closer to $0.11–$0.13.

Retell publishes a range — $0.07 to $0.31 a minute — and the range is the honest part. It's a range because you assemble the stack:

• Retell Voice Infrastructure — $0.055/min.Orchestration plus speech-to-text. This is the piece you can't opt out of.
• Text-to-speech — $0.015/min for Retell, MiniMax, Fish, Cartesia, OpenAI and Inworld voices. $0.040/min if you want ElevenLabs voices. Remember that number.
• Telephony — $0.015/min, varying by country, and $0 if you bring your own SIP trunk.
• LLM — $0.0016/min on GPT 5 nano up to $0.064/min on GPT 5.6 Terra or Claude 5 Sonnet. A 40x spread, and it is a dropdown.

Add the floor of every line and you get $0.0866 a minute. Add the ceiling of the model line and you get $0.149. That is the same $0.07–$0.31 range, just itemized — which is the part that lets you actually manage it.

I run a Retell agent in production on the MiniMax voice tier, with post-call webhooks piped through an n8n workflow that drops a summary into Telegram. The thing I check first on the invoice is never the infrastructure line. It's the model line, and I'll show you why in a minute.

Every comparison post you'll find argues about per-minute rates. Almost none of them mention the number that will actually break your deployment, which is how many calls the platform will let you run at the same time.

On ElevenLabs, concurrency is welded to the subscription tier.You want 30 simultaneous lines, you buy Scale at $299 — and 3,738 minutes you may not need. You want 40, you buy Business at $990. There is no way to buy the ceiling without buying the minutes.

Burst pricing exists to soften that, and it's worth reading the burst pricing docs carefully, because all three clauses matter:

• Burst accepts calls up to 3x your concurrency limit, or 300, whichever is lower.
• Those calls bill at 2x the normal rate.
• Burst calls get lower priority for speech-to-text and text-to-speech, so they can run slower than your normal calls.
• Past the burst ceiling, calls are rejected with an error unless you've enabled call queueing on the agent.

Read that as an operator rather than as a buyer. Your busiest hour — the one where the calls are worth the most — is precisely the hour when your overflow calls cost double and run on deprioritized speech processing. You are paying a premium for your worst-sounding calls.

On Retell, concurrency is a line item. Twenty concurrent calls come with pay-as-you-go, and additional slots are $8 per concurrency per month. Enterprise removes the cap entirely. Getting to 40 lines costs you $160 a month on top of usage, against a $990 plan on the other side.

This is a normal shape for a busy clinic, a roofing company, or an agency running reception for a client. Three thousand billable minutes a month, peaking at twenty-five simultaneous calls on a Monday morning.

A note on the ElevenLabs column, because the cheap-looking path is a trap. Pro at $99 plus 1,762 overage minutes at $0.08 comes to $240, which is about $60 lessthan Scale. But Pro caps you at 20 concurrent calls and your peak is 25. Those five overflow calls bill at $0.16 and run deprioritized. The arithmetic is telling you to save sixty dollars by degrading your busiest hour. Don't take that trade.

Look at the table again and read it sideways instead of down.

Changing platform moved the bill about $100 a month. Changing the model moved it $187.On the same volume, on the same platform, with the same voice. That is the single most useful thing in this comparison, and it is not a fact about either vendor — it is a fact about voice agents.

The reason is that a voice agent's LLM turn is short. You are not sending 40,000 tokens of context; you are sending a system prompt, a handful of turns, and a tool result. Frontier reasoning buys you almost nothing in that shape, and you pay 40x for it. Most of what makes a phone agent feel smart is prompt structure, tool wiring, and turn-taking tuning — not model size.

Which is also why the itemized bill is worth something on its own. On Retell the model is a dropdown, so the experiment costs you an afternoon. On ElevenLabs the LLM is an external bill you assemble yourself, which is more control in theory and more places to lose track of spend in practice. If you're already fighting model costs elsewhere, the same logic that drives routing cheap tasks to cheap models applies here, with sharper edges, because every second of silence is still billing.

This is the part that collapses most of the debate. The usual argument for ElevenLabs Agents is voice quality, and it is a fair argument — their voices are the benchmark.

But Retell sells those same voices as a TTS option at $0.040 a minute against $0.015 for its other tiers. That is a $0.025 premium, or $75 a month at 3,000 minutes. Add it to the Retell column and you land at $375 — still under the $401 ElevenLabs Scale total, with the same voice coming out of the speaker.

The voice is a purchasable component. The platform is a commitment.Don't buy the second to get the first.

I'd rather give you the case where my own conclusion weakens than pretend it doesn't exist. Run the Business-tier shape: 12,375 minutes, 40 concurrent lines.

• ElevenLabs Business: $990 plan, plus roughly $248 of LLM and $173 of telephony → about $1,411.
• Retell on a cheap model: $0.0866/min across all lines is $1,072, plus $160 of extra concurrency → about $1,232.
• Retell on Claude 5 Sonnet: $0.149/min is $1,844, plus $160 → about $2,004.

So at real volume the ranking holds only while your model is cheap. Put a frontier model on every call and ElevenLabs' bundled-minutes plan wins on price, because you've stopped paying per-minute for the expensive part and started paying per-token for it instead. Run the numbers with your model, not with mine.

And budget for the things neither pricing page shows you: silence still bills, failed calls still bill, and every voice platform I've deployed has had a gap between the demo and the deployment.

Strip out the marketing and there are four questions that settle it.

1. Is your traffic spiky? If your peak is more than about 1.5x your average concurrency, go Retell. Buying concurrency at $8 a slot beats buying a plan tier, and you avoid burst-rate calls landing in your busiest hour.
2. Do you need to change models later?On Retell that's a dropdown. On ElevenLabs it's your own provider integration. If you expect to tune cost down after launch, that difference compounds.
3. Do you need HIPAA and a BAA? Retell puts PII redaction on pay-as-you-go and HIPAA/BAA on Enterprise. On ElevenLabs, compliance sits behind Enterprise too. Get it in writing before you build either way.
4. Is the voice the product? If you're selling a persona — a character, a branded concierge, something where the voice isthe differentiator — and you're already inside ElevenLabs, staying there is a reasonable call. For a receptionist that books appointments, it is not.

For the work I actually get hired for — inbound booking, lead qualification, after-hours reception for local service businesses — it's Retell, on a cheap model, with a $0.015 voice, and I spend the savings on turn-taking tuning instead. If you're still choosing between orchestration layers rather than between these two, the Retell vs Vapi breakdown is the other half of this decision.