The cheapest AI voice agent in 2026: published per-minute prices, compared

Among managed AI voice-agent platforms with published, computable pricing, the lowest complete per-minute engine rate we found is ThunderPhone Spark at 2¢ per minute; speech recognition, the language model, and voice are included in one pay-as-you-go rate. As illustrative planning arithmetic, not a quote, lean DIY component stacks compute to roughly the same floor before engineering work, but this conclusion is limited to prices published and checked on September 1, 2026; call length, language, telephony, and configuration can reorder the results, so rerun the arithmetic for your workload.

“Complete engine rate” matters here. Spark covers the voice-AI engine, not every possible cost of placing a phone call. Telephony depends on setup, selected premium voice or language paths can add up to 3¢ per minute, and optional features have published surcharges. The claim is therefore narrower, and more useful, than saying one product is always cheapest.

Why “cheapest voice AI” claims are slippery

A per-minute headline may describe an entire engine, one infrastructure component, a hosting fee before provider bills, or an overage after a monthly allowance. Other services bill per call or per unique customer, making direct per-minute comparisons impossible without assumptions. Enterprise products may publish only a contract minimum.

Those differences are large enough to reverse a ranking. Vapi’s $0.05 per minute is hosting before speech, model, voice, and telephony costs, while Telnyx’s $0.05 covers orchestration, speech-to-text, and text-to-speech but leaves model tokens and telephony on top. Vapi pricing and Telnyx pricing, accessed 2026-09-01. A subscription can look inexpensive until overage dominates, while a higher per-minute tier can still cost more after its platform fee.

For the deeper taxonomy, see how voice AI platforms price and the AI phone-agent cost guide. The table below keeps each vendor’s billing unit intact before normalizing one scenario.

Published pricing at a glance

Phonely’s discrepancy deserves explicit treatment: its plan cards say Starter includes 200 minutes and Professional includes 650, while the comparison table on the same page says 250 and 750. Its FAQ also says Starter supports up to 50 calls while the card estimates about 100. Confirm the live allowance before buying. Phonely pricing, accessed 2026-09-01.

Worked example: 10,000 connected minutes

Every computation in this section is illustrative planning arithmetic, not a quote. Use the corpus-standard workload: 2,000 calls × 5 connected minutes = 10,000 connected minutes in a month. Assume one primary included language, no transfers or optional add-ons, one phone number where relevant, and traffic distributed evenly enough to fit the stated tier limits.

• ThunderPhone Spark: 10,000 × $0.02 = $200 for the engine. A selected premium voice or language path at the maximum surcharge would add up to 10,000 × $0.03 = $300; telephony remains setup-dependent.
• Telnyx: using Telnyx’s own realistic $0.056/minute example, 10,000 × $0.056 = about $560. Telnyx pricing, accessed 2026-09-01.
• Retell AI: applying its advertised range gives 10,000 × $0.07 = $700 through 10,000 × $0.31 = $3,100. The selected model and add-ons decide where a configuration lands. Retell pricing, accessed 2026-09-01.
• Vapi: hosting alone is 10,000 × $0.05 = $500, before speech, model, voice, and telephony costs. A complete total cannot be computed from the public page. Vapi pricing, accessed 2026-09-01.
• Bland AI: Start is 10,000 × $0.14 + $0 = $1,400; Build is 10,000 × $0.12 + $299 = $1,499; Scale is 10,000 × $0.11 + $499 = $1,599. Start is cheapest only if the 2,000 monthly calls stay within its 100-calls-per-day and 10-concurrent-call limits. Bland pricing, accessed 2026-09-01.
• Phonely Starter on monthly billing: using the comparison table’s 250 included minutes, $50 + (10,000 − 250) × $0.25 = $2,487.50. Using the plan card’s newer 200-minute figure, $50 + (10,000 − 200) × $0.25 = $2,500. The source disagrees with itself, so treat this as a range and confirm the allowance. Phonely pricing, accessed 2026-09-01.
• Smith.ai AI Receptionist: the five-minute call assumption gives 10,000 ÷ 5 = 2,000 calls. On the published 500-call Enterprise option, $800 + (2,000 − 500) × $2.10 = $3,950. This conversion depends completely on average call length. Smith.ai AI Receptionist pricing, accessed 2026-09-01.
• Goodcall: minutes do not determine the bill. If those 2,000 calls come from 1,500 unique customers, Scale computes as $249 + (1,500 − 500) × $0.50 = $749. Fewer repeat callers would cost more; more repeat calls from the same people would not add unique customers. Goodcall pricing, accessed 2026-09-01.
• Synthflow: the published $30,000 annual contract floor does not disclose how many minutes it buys, so this workload has no public total. Synthflow pricing, accessed 2026-09-01.

In this defined scenario, ThunderPhone Spark’s published engine arithmetic is lowest among the managed platforms with comparable, computable usage prices. That is not a universal ranking. Different languages, call lengths, call distributions, configurations, and telephony arrangements can change both eligibility and ordering.

What the cheapest DIY voice AI actually costs

The attention-grabbing DIY floor is real, but it needs attribution. The open-source hypercheap-voiceAI repository claims costs “as low as $0.28 per hour ($0.0046 per minute)” and a typical $0.25–$0.35 per session hour. Those are third-party project claims, not verified vendor list prices, and some underlying speech-recognition figures could not be independently checked. hypercheap-voiceAI, accessed 2026-09-01.

Public component prices establish a more reproducible floor. Streaming speech-to-text has list-price reference points around $0.0025–$0.006 per minute: an illustrative conversion, not a quote, is AssemblyAI’s $0.15/hour ÷ 60 = $0.0025/minute, while OpenAI lists $0.003–$0.006/minute and Deepgram lists $0.0048 promotional and $0.0077 regular pricing. AssemblyAI pricing, OpenAI API pricing, and Deepgram pricing, accessed 2026-09-01.

LLM inference adds roughly $0.001–$0.01 per conversation minute as an illustrative planning estimate, not a quote, depending on model choice and context growth. That range rests on published token prices and conversation-token assumptions, not a flat per-minute tariff. OpenAI API pricing, Google Gemini API pricing, Together AI pricing, and Inworld’s voice-agent cost model, accessed 2026-09-01.

Text-to-speech list prices around $4–$30 per million characters convert, at about 1,000 generated characters per speaking minute, to an illustrative $0.004–$0.03 per generated-speech minute, not a quote; only the agent’s speaking share is billed. Google Cloud TTS pricing, Amazon Polly pricing, and Inworld TTS pricing, accessed 2026-09-01. Hosted orchestration can add $0.01/minute, while US telephony reference points run from $0.0032/minute inbound on Telnyx to $0.018/minute PSTN on Daily/Pipecat Cloud. LiveKit pricing, Telnyx SIP pricing, and Daily/Pipecat pricing, accessed 2026-09-01.

One lean-stack illustration makes the floor concrete: $0.0025 STT + $0.001 estimated LLM + ($0.004 TTS × 50% speaking share) + $0.01 orchestration + $0.0032 telephony = $0.0187 per conversation minute, or 1.87¢. This is illustrative planning arithmetic, not a quote, and it mixes category-floor assumptions that may not meet a particular quality, latency, reliability, or geography requirement. Lean combinations therefore land around 1–4¢ per minute before engineering. AssemblyAI pricing, OpenAI API pricing, Google Cloud TTS pricing, LiveKit pricing, and Telnyx SIP pricing, accessed 2026-09-01.

The missing line items are engineering time, monitoring, redundancy, concurrency headroom, incident response, compliance work, and support. A component demo and a production phone system are different deliverables. See the full voice AI cost breakdown and ThunderPhone versus building your own before treating raw API spend as total cost.

How a managed 2¢-per-minute rate compares

ThunderPhone Spark’s 2¢ engine rate sits inside the illustrative DIY component-floor band without requiring the buyer to assemble speech recognition, a language model, and voice services. It is prepaid and pay as you go, with no subscription, seat fee, or platform fee. Signing up and building are free; making real calls requires a payment method and at least $3.00 in prepaid balance, and funds do not expire.

The boundary is equally important. Telephony is not part of the engine rate and depends on whether a customer uses an imported VoIP/SIP number or a ThunderPhone demo number. A selected premium voice or language path can add up to 3¢ per minute, and optional features such as supervision, long prompts, demo-number usage, or live transcripts can add to the configured rate. The builder shows the all-in per-minute engine rate for the exact configuration before calls run, while Billing History itemizes calls afterward.

At very low volume, a flat subscription with bundled numbers may be simpler even when its unit economics are higher. Phonely Starter, for example, bundles multiple numbers with its monthly plan, subject to the live-page allowance discrepancy noted above. At high volume, lower rates attached to platform fees or custom commitments can amortize differently. Price convenience, operational limits, and total invoices, not just the smallest headline.

FAQ

What is the cheapest AI voice agent platform?

Among the managed platforms with published, computable pricing reviewed here, ThunderPhone Spark has the lowest complete base-engine rate at 2¢ per minute. That rate includes speech recognition, the language model, and voice. Telephony and selected options can add cost, and changes in language, configuration, or call pattern can reorder the comparison.

Can you really run an AI voice agent for 2 cents a minute?

Yes, ThunderPhone publishes a 2¢-per-minute Spark engine rate, and DIY component arithmetic can approach the same level. Spark includes the three core engine components in one rate; it does not include setup-dependent telephony or every optional add-on. The lean DIY example above computes to 1.87¢ per conversation minute, but it is illustrative planning arithmetic, not a quote, and excludes engineering and operations.

Is building your own voice agent cheaper than a platform?

It can be cheaper in raw API spend, but not necessarily in total cost. A community repository claims as little as $0.0046 per minute, while reproducible vendor list-price combinations suggest a lean 1–4¢ component floor. Engineering, monitoring, redundancy, concurrency, compliance, and support sit outside that number. hypercheap-voiceAI, AssemblyAI pricing, Google Cloud TTS pricing, LiveKit pricing, and Telnyx SIP pricing, accessed 2026-09-01.

What hidden costs make “cheap” voice AI expensive?

Add-on stacking, overage, platform fees, concurrency, telephony, and compliance can outweigh the headline. Retell lists several $0.005/minute add-ons, $0.01/minute PII removal, and $0.10/minute AI quality assurance after its free allowance; added concurrency is $8 per concurrent call monthly. Retell pricing, accessed 2026-09-01. Phonely displays Starter overage at $0.25/minute monthly and $0.35/minute under annual billing. Phonely pricing, accessed 2026-09-01. Bland’s Build and Scale platform fees are $299 and $499 monthly. Bland pricing, accessed 2026-09-01. Vapi lists HIPAA at $2,000/month and zero data retention at $1,000/month. Vapi pricing, accessed 2026-09-01.

How much does 10,000 minutes a month cost?

In the defined scenario, published arithmetic gives ThunderPhone Spark $200; Telnyx about $560; Retell $700–$3,100; Bland Start $1,400 if its limits fit; and Vapi $500 for hosting alone before provider and telephony costs. Phonely Starter computes to $2,487.50–$2,500 because its page disagrees on included minutes; Smith.ai computes to $3,950 under the five-minute-call assumption; Goodcall computes to $749 under the 1,500-unique-customer assumption. These are illustrative calculations, not quotes. Telnyx pricing, Retell pricing, Bland pricing, Vapi pricing, Phonely pricing, Smith.ai AI Receptionist pricing, and Goodcall pricing, accessed 2026-09-01.

Sources and freshness

All third-party figures were checked against the linked first-party vendor pages on September 1, 2026. The hypercheap-voiceAI numbers are explicitly labeled as third-party repository claims. This comparison is documentation-based; no vendor accounts, invoices, or live calls were tested. Prices and plan terms change, so verify the live pages and rerun the arithmetic for your own call length, language, telephony, and configuration before relying on any result.