Ask five people building a voice agent which vendor comes to mind first, and four of them will say ElevenLabs. That's not an accident. Their TTS models set the bar the rest of the industry gets measured against, and "sounds like ElevenLabs" has become shorthand for "sounds real." None of what follows is an argument against their voice quality. It's an argument for reading past the name before you build an entire voice agent around it.
The pitch is one platform. The bill is three.
ElevenLabs sells its Agents Platform as an all-in-one way to ship a voice agent: speech in, speech out, orchestration in between. Read their own pricing documentation and a different picture shows up. LLM tokens get billed separately, at the model provider's own rates, deducted from your ElevenLabs account balance rather than folded into the per-minute agent price. Telephony is billed separately too. Twilio, or whatever SIP carrier you route calls through, charges you directly, on top of everything else. So the "one platform" under evaluation is really three vendor relationships wearing one login screen: ElevenLabs for voice, an LLM provider for reasoning, a telephony carrier for the phone call itself. That's a completely normal way to architect a voice stack. It's just not what "all-in-one" usually implies, and it's easy to miss until you're three invoices in.
Usage-based pricing, gated by a plan tier
The per-minute agent rate looks like straightforward usage-based pricing: pay for what you use. How much you can use at once, though, is capped by which subscription tier you're on, not by how much you're willing to spend. If your agent needs to field 25 simultaneous calls, no amount of pay-as-you-go spending on a mid tier gets you there, a specific higher plan is required regardless of actual per-minute usage that month. That's a hard ceiling hiding inside what reads, at first glance, like pure usage-based pricing.
Concurrency is a plan-tier ceiling, not a usage dial
Included minutes (bars) vs. max concurrent calls (line), by plan tier
Going over your included minutes has its own wrinkle. Standard overage runs a flat $0.08 per minute, but ElevenLabs' own pricing page lists a separate "burst" rate of $0.16, exactly double, for load that spikes past a plan's steady-state capacity. Voice agents are exactly the kind of product that spikes: a marketing push, a support outage, a slow news day turning into a busy one. Budgeting for the standard rate and getting billed at the burst rate is an easy mistake to make once, and an expensive one to make twice.
Four different billing units, one final invoice
Add it all up, and pricing a real ElevenLabs voice agent means converting across at least four separate billing units before even reaching the separately-billed LLM and telephony legs: credits per character for the base TTS models, a flat rate per minute for the Agents product, credits per minute for adjacent features like voice changing or dubbing, and dollars per hour for Scribe transcription, at two different rates depending on whether it's batch or realtime. None of these numbers are hidden. They're all published, and every one of them checks out against ElevenLabs' own pricing and documentation pages. But nobody hands over a single formula, and a platform this many products deep doesn't compress into one number by accident.
Why this matters more than it sounds like it should
None of this makes ElevenLabs a bad choice. Their voice models are genuinely excellent, and for a lot of use cases the brand recognition alone is worth something: an investor, a partner, or a user who's already heard of ElevenLabs is one less thing to explain. But "the name I already know" and "the right architecture for my call volume, latency requirements, and budget" are two different questions. The first one is a lot easier to answer by reflex than the second one is to answer by actually reading a pricing page closely enough to notice the concurrency caps and the burst rate.
That's the evaluation we run before recommending or building on any voice stack at Dissolve Labs: read the actual docs and pricing tables line by line, benchmark latency and cost across every leg of the pipeline instead of just the TTS one, and size the plan tier against real concurrency needs before a launch finds the ceiling first. If you're about to default to the first name that comes to mind for your voice agent, it's worth being the second thing checked, not the only one.