ElevenLabs Voice Agents: Price Approved Audio, Then Add the Phone Bill
Voice generation is one product. A live voice agent is another: LLM, telephony, concurrency and retakes. August’s agent enthusiasm will push people to conflate them.
This page is periodically reviewed to reflect current pricing and plan changes.
Fastest win
If you need a voiceover, buy voiceover. If you need a phone agent, budget the LLM and the call path separately. Mixing them on one “AI voice” line is how this category overruns.
Two products sharing a brand
ElevenLabs will happily sell you both narration and agents. They are not the same accepted outcome. Narration: approved minutes in a timeline. Agent: a completed call that followed policy, at a concurrency you actually hit.
Agent-safety news from the text-agent world applies here too. A voice agent that can act is an agent with a mouth. Permissions still matter. A pretty voice does not reduce refund risk.
Cost per approved minute versus cost per completed call
For voiceover, count plan + retakes + mastering. A cheap generator that mis-says your product name ten times is not cheap.
For agents, count voice + LLM + telephony + failed dialogs + human takeover. If you only quote ElevenLabs’ character rate, you are missing most of the bill.
Confirm current ElevenLabs plan matrices and API tables on their pricing page. They change by product surface.
ElevenLabs for generated voice; Descript when editing is the job
ElevenLabs is still the specialist to test first for generated voice, dubbing and agents. Descript wins when the labour is editing recorded media, not synthesising it.
Rights and cloning
A cloned voice without a documented consent trail is not a cost optimisation. It is a legal and reputational cost waiting for a date.
If you need a house voice, budget the rights work as part of the unit cost. If you cannot get consent, use a stock or platform voice and stop pretending cloning is “just a feature”.
This is one of those cases where the cheapest audio file is not the accepted outcome.
Ranked recommendation
Best choice for narration and dubbing: ElevenLabs, smallest plan that covers accepted minutes, after a pronunciation test.
Best choice when the job is editing real recordings: Descript.
Best choice for a support voice agent: only after a written tool allowlist and a telephony estimate; start with a low concurrency cap.
Avoid cloning a voice without a rights trail. Avoid planning as if voice minutes never run out.
Key Takeaways
- →Voiceover and voice agents are different denominators.
- →Retakes and pronunciation repair belong in the numerator.
- →Telephony and LLM cost sit outside the pretty per-character rate.
- →Descript is the editor; ElevenLabs is the specialist generator.
Editorial context
Who is this for?
Podcasters, marketers and support leads adding “voice AI” after the agent news cycle.
When NOT to use this
People who need on-camera avatars. That is HeyGen or Synthesia. Different minute, different rights problem.
Pricing insights
Characters, minutes, credits, concurrency and telephony do not add up on a napkin. Split narration jobs from live-agent jobs in the model.
Alternatives to consider
Murf or LOVO for simpler recurring voiceover. Descript for interview cleanup. Native OS or ChatGPT voice for casual use.
Final verdict
Test ElevenLabs first for generated voice. Do not fund a voice agent until the call path and LLM are priced. Measure approved minutes and pronunciation repair.
Frequently Asked Questions
Why is ElevenLabs voice-agent pricing hard to compare?
Characters, minutes, credits, concurrency and retakes are different meters. Convert them to approved voice tracks or resolved voice-agent turns after legal review.
When is Descript the better buy?
When the job is edit-heavy podcast or interview work, not a branded voice agent. Do not pay agent rates for a timeline you will cut by hand anyway.
What should we confirm before cloning a voice?
Rights, consent and commercial-use terms on the current vendor page. A cheap clone that legal rejects is not an accepted outcome.