Free AI · Updated August 24, 2026 · 8 min read

Build a Voice AI Agent for Free in 2026: The Whole Stack

Every layer of a production voice agent — transport, speech-to-text, LLM, text-to-speech — has a real free tier in 2026. What each gives you and where the ceiling actually is.

A voice agent has four moving parts — transport, speech-to-text, a model, and text-to-speech — and in 2026 every one of them has a free tier substantial enough to get a working agent onto a real phone number without a card. This is the current stack, what each layer actually gives you, and, more usefully, which limit you will hit first.

The stack, and what each layer costs

LayerFree optionWhat you get free
Transport / orchestrationLiveKit1,000 agent-session minutes, 5,000 WebRTC minutes, 1,000 SIP minutes, 1 phone number, $2.50 inference credit — per month
Speech-to-textDeepgram$200 signup credit, no card
Speech-to-text (alt)AssemblyAI$50 signup credit
LLMGroq or CerebrasRecurring daily free allowance, no card
Text-to-speechElevenLabsRecurring monthly character allowance
Text-to-speech (alt)CartesiaStartup grant track, application required

The two shapes matter. Signup credits (Deepgram, AssemblyAI) are one-time dollar grants — spend them and they are gone. Free tiers (LiveKit, Groq, Cerebras, ElevenLabs) reset, so they support an agent running indefinitely at low volume. A prototype wants the credits; anything you leave running wants the tiers.

The limit you will actually hit first

Not the minutes. LiveKit's free plan caps you at 5 concurrent sessions and 1 agent deployment.

This is the single most common surprise, because the number that looks binding — 1,000 agent-session minutes a month — is roughly 16 hours, which sounds like plenty. And it is, for a demo. But a demo with six people on it fails immediately, and it fails in the most visible way possible: in front of the six people you invited.

So plan around concurrency, not volume:

  • Solo development and recorded demos: comfortably free.
  • A live demo to a small group: fine up to 5 simultaneous callers, and not one more.
  • Anything with real users: you are on a paid plan the day you launch, regardless of your minutes.

The $2.50 monthly inference credit is for LiveKit's own bundled model routing. If you bring your own provider key — which you will, to use the Groq or Cerebras free tier — it does not get consumed, so ignore it in your planning.

Choosing the speech-to-text layer

Deepgram's $200 signup credit is the largest no-card grant in this stack by a wide margin, and it is the right default. At streaming transcription rates that is a large number of hours of audio — enough to run a prototype well past the point where you know whether the product works.

AssemblyAI's $50 is smaller but real, and worth having as the second key you can fail over to. Both are self-serve with no application.

The trap in both cases is that they are one-time. Unlike the LLM tiers, they do not come back. So do your throwaway testing — the loops where you are debugging audio format issues and burning minutes on silence — against local Whisper before you point anything at a paid endpoint. Cheapest speech-to-text API has the per-hour comparison for when the credits run out, and Whisper API cheap covers the self-hosted route.

Choosing the model layer

For voice, latency dominates quality far more than it does in a chat product. A 400ms difference in first-token latency is invisible in a chat UI and painfully obvious in a conversation, where humans expect a response gap under about 500ms.

That makes Groq and Cerebras unusually well-suited here: both are hardware-first inference plays whose entire proposition is tokens-per-second, and both have recurring free tiers with no card. Groq vs Cerebras compares them directly — the short version is that Cerebras currently wins the raw speed benchmarks on large models and has the more generous free tier, while Groq wins on model breadth.

Use a small, fast model for the conversational turn and reserve anything larger for the occasional request that genuinely needs reasoning. In a voice agent the user is waiting in real time, so the usual advice about routing by difficulty is worth more than it is anywhere else.

Choosing the text-to-speech layer

ElevenLabs has the recurring free character allowance and the voice quality most people are benchmarking against. It is the default.

Cartesia runs a startup grant track, which is an application rather than a free tier, but it is worth applying to if you are past the prototype stage — the latency profile is built for exactly this use case. Cheapest TTS API covers the per-character economics once you are paying.

What this stack costs when you outgrow free

Being honest about the cliff, because every layer has one:

  • LiveKit: the day you exceed 5 concurrent sessions. This is a launch-day cost, not a scale cost.
  • Deepgram / AssemblyAI: when the one-time credit is spent. There is no recurring free tier underneath it.
  • Groq / Cerebras: rate limits rather than a hard stop — you degrade rather than fail, which is the friendliest cliff in the stack.
  • ElevenLabs: when the monthly character allowance runs out, which on a chatty agent is sooner than you expect. Characters, not requests.

The realistic read: the free stack takes you from nothing to a working agent on a real phone number, and through as much iteration as you want with a handful of testers. It does not take you to production with real users. That is a fair trade, and it is a much better starting position than the same question would have had in 2024.

After the free tiers

Once you are past prototyping, the startup programs are where the numbers get large: Deepgram's startup program runs to $100,000 over twelve months, and AssemblyAI's is one of the largest credit pools available to an early voice company. Both require an application and a real product rather than a prototype, which is the right order to do this in anyway.

Perkstack tracks all of these with the date each was last verified against the provider's own page. Create a free account to see the full stack with claim guides, or browse the catalog.

rest of this guide

The rest of this guide picks up at "Choosing the model layer".

  • Choosing the model layer
  • Choosing the text-to-speech layer
  • What this stack costs when you outgrow free
Create a free account

Free account. No card. 230+ verified perks.

Not ready for an account? Find out what you can claim first.

Already have an account? Log in

Frequently asked questions

Can you really build a voice AI agent for free?

Yes, through prototyping. Transport (LiveKit), speech-to-text (Deepgram's $200 credit), the model (Groq or Cerebras free tiers) and text-to-speech (ElevenLabs) all have real free options, and LiveKit's free plan even includes a phone number. It does not reach production with real users.

What is the first limit you hit on a free voice agent stack?

LiveKit's cap of 5 concurrent sessions, not the minute allowances. A thousand agent-session minutes a month is about 16 hours, which is plenty for development — but a live demo with six simultaneous callers fails immediately.

Which free speech-to-text credit is biggest?

Deepgram's $200 signup credit, with no card required. AssemblyAI's $50 is the next largest. Both are one-time grants rather than recurring tiers, so they do not come back once spent.

Which model should a voice agent use?

A fast one. Latency matters far more in voice than in chat because humans expect a response gap under about half a second. Groq and Cerebras are both built for tokens-per-second and both have recurring free tiers with no card.

the weekly drop

One email, Sundays.

New verified perks and the provider prices that moved. Free, unsubscribe in one click.

Keep reading

Building on AI? Don't pay full price.

Perkstack tracks 230+ verified AI credits, free signup credits and startup grants, each with its verified conditions and apply link.

the weekly drop

One email, Sundays.

New verified perks and the provider prices that moved. Free, unsubscribe in one click.