Voice AI Price Index (September 2026): published per-minute prices for 15 platforms, what each includes, and a computable floor

This index uses first-party pricing pages only, accessed 2026-09-10 except for four rows last checked on 2026-09-01. The method targets the cheapest published US-inbound configuration, flags missing components, and separates usage floors from fixed charges. The reference workload is 2,000 calls × 5 connected minutes = 10,000 connected minutes/month: one included language, one number, no transfers/hold/voicemail, and ≤10 concurrent calls.

Published prices and computable floors

How to read the floor column

These figures are illustrative planning arithmetic, not a quote. A floor measures the listed usage configuration; it is not an invoice estimate or a quality ranking. STT means speech-to-text, TTS means text-to-speech, and LLM means language model.

1. Published-only arithmetic. Use published first-party rates, without negotiated discounts or unsupported assumptions about bundled services. Preserve uncertainty where the page leaves a component unclear. “Not computable” means the available evidence cannot support a complete comparable rate; it does not mean the service is expensive. Provider combinations are estimates, not vendor offers.
2. US inbound local PSTN. Add the published local inbound phone leg and required streaming fees for the stated route. Exclude toll-free and outbound rates. Keep number rental, fixed subscriptions, and recording outside the per-minute floor. Add applicable fixed charges when estimating a monthly bill. A browser session or bare SIP connection is not automatically a complete PSTN call.
3. Cheapest listed LLM, or an explicit exclusion. Use the cheapest listed compatible option where its price is computable. Mark external LLM costs separately instead of treating them as free. Telnyx uses its published example's token assumption, not a guaranteed model minimum. Deepgram uses bundled Standard PAYG, while Cartesia's temporary inclusion applies specifically to UI-created agents.

The floor column follows the published configurations, rather than optimizing every possible carrier and subscription combination. Deepgram's commitment pricing and Bland's paid tier can lower a marginal rate while changing fixed costs. ElevenLabs' overage rate is not its average subscription cost. Read the pricing model and fee columns together before comparing rows.

Callplane's estimated subtotal adds its platform charge, Twilio inbound and streaming, and Deepgram transcription; TTS and LLM remain excluded. Its managed Cloud provider rates are not itemized, so the BYO estimate cannot establish a complete Cloud price. The conditional subtotal should travel with that qualification whenever it is cited.

Bland's Start configuration also depends on daily traffic distribution. A monthly call count alone does not establish that every day fits the plan. Likewise, the reference workload's concurrency assumption is a ceiling, not a promise that every free plan supports it. Cartesia's plan eligibility and concurrency need checking before applying its conditional floor.

ThunderPhone's engine includes the model but excludes carrier charges. A Plivo customer can retain Plivo telephony and connect the agent through manual SIP. The published combination is $0.02/min Spark + $0.0055/min Plivo inbound local, plus $0.50/month for the Plivo number. This assumes that inbound route without additional billable legs. ThunderPhone pricing and Plivo US pricing, accessed 2026-09-10. Telnyx numbers can also connect through ThunderPhone's Telnyx or manual SIP connection; carrier charges remain separate.

Billing clocks can also differ from connected conversation time. Deepgram meters websocket connection time. Its floor therefore assumes the billed connection tracks the call duration. Unverified rounding rules are not silently normalized across platforms. Exact monthly totals require the relevant billing increments, allowances, traffic distribution, and configuration.

For buying context, read the cheapest AI voice-agent guide alongside this dated reference. The pricing-model comparison explains bundled engines, component billing, and subscriptions. The AI phone-agent cost guide and cost breakdown cover budgeting; ThunderPhone pricing explained details its configuration charges.

Download the data

The same rows are available as JSON and CSV, with one record per platform and the same fields as the table (pricing model, headline rate, inclusions, what bills on top, computable floor, platform fee, source URL, access date).

License: CC BY 4.0, attribution: ThunderPhone Voice AI Price Index. Send corrections to alex@thunderphone.com.

When citing a row, retain its access date, billing unit, exclusions, and floor qualification. Copying a hosting fee into an “all-in price” field changes the meaning. Missing costs are unknown, not zero. Preserve the promotional cutoff when reusing Deepgram's figures, and the temporary-offer qualification when reusing Cartesia's.

Index changes

• 2026-09-10: Initial publication. The table establishes the dated baseline. Phonely, Smith.ai AI Receptionist, Goodcall, and Synthflow retain their 2026-09-01 checks; they were not reverified with the newer entries.
• 2026-09-10: Bland tier change recorded. Scale's $0.11/min rate and $499/month fee no longer appear on its pricing page. They are historical figures, not current offers. Bland pricing, accessed 2026-09-10.
• 2026-09-12: Deepgram promotion ends. Standard PAYG moves from $0.056/min through that date to $0.075/min from 2026-09-13. Both periods remain visible in the row. Deepgram pricing, accessed 2026-09-10.
• 2026-09-10: Cartesia promotion recorded. LLM usage is free for UI-created agents for a limited time. An end date is not published on the page. Cartesia pricing, accessed 2026-09-10.

FAQ

How often is the index updated?

This is a dated snapshot; no fixed refresh schedule is promised. Use each row's access date and the change log to establish freshness. A newer page-level date does not make the older retained rows newly verified.

Why do floors differ from headline rates?

Headlines cover different layers of a call. A floor adds the documented components for the stated inbound route. Unknown LLM costs, conditional bundles, and fixed fees remain visible instead of becoming unsupported complete totals.

Why is ThunderPhone in the index it publishes?

ThunderPhone publishes this index and sells a product included in it. That commercial interest is disclosed here. Its row uses the same inclusion, exclusion, source, and date fields. Its engine price does not become a complete phone-call price merely because ThunderPhone publishes the table.

How do I submit a correction?

Email alex@thunderphone.com with the affected row and a first-party source. Include the plan, billing unit, access date, and the specific figure or inclusion that changed. An account-specific quote can explain your experience, but cannot establish a generally available public floor.

Sources and freshness

First-party sources, accessed 2026-09-10:

• ThunderPhone pricing.
• Plivo agent pricing and US voice pricing.
• Telnyx Voice AI pricing.
• Deepgram pricing.
• Cartesia pricing.
• Twilio ConversationRelay pricing, US voice pricing, and Media Streams pricing.
• Callplane pricing.
• Retell pricing.
• Vapi pricing.
• Bland pricing.
• ElevenLabs Agents pricing.

Retained first-party sources, accessed 2026-09-01:

• Phonely pricing.
• Smith.ai AI Receptionist pricing.
• Goodcall pricing.
• Synthflow pricing.

This page is documentation-based; no account, invoice, or live call was tested. Access dates record when each linked page was last checked.