Retell AI vs Vapi vs Bland AI: Real Voice Agent Cost per 10,000 Minutes in 2026

The short answer: the headline per-minute rate on every voice-agent platform is one line of a four-line bill, and it is usually the smallest one. Vapi's platform fee is $0.05 a minute, Retell's is $0.055, and both pass the LLM, text-to-speech, speech-to-text and telephony costs through separately; Bland bundles the model layers into one rate ($0.11–$0.14 depending on plan) but still bills telephony on the side. Once every layer is added, a support agent that uses tools and retrieval costs roughly $0.167–$0.172 per minute on Vapi and Retell against $0.123 per minute of variable usage on Bland Scale, and an outbound sales agent with a heavier model climbs to $0.247–$0.252 per minute on the bring-your-own-model platforms while Bland's bundled variable rate stays flat. Bland's $0.123 is usage only: Scale also carries a $499 monthly plan fee on top, so a full month at 10,000 minutes costs $1,729 on Scale, not the $1,230 that the variable rate alone implies, and Scale's plan is capped at 5,000 calls a month, a limit that matters as much as the per-minute rate when choosing a Bland tier. Platform minutes and telephony minutes are not the same meter, and neither is billed the same as the underlying AI minutes the model actually consumes.

What each vendor bills, and for what

Retell AI (REPORTED, from Retell's own pricing page as read by third parties, August 2026). Three separate tables on one pricing page: a Conversation Voice Engine at $0.055 a minute (the platform fee), a text-to-speech table from $0.015 to $0.040 a minute depending on voice provider, and an LLM table where a frontier model (GPT-5.5-class) costs $0.16 a minute standard and $0.32 a minute on a fast tier, roughly three times the platform fee by itself. Concurrency is capped at 20 concurrent calls on pay-as-you-go, with additional capacity at $8 per concurrent call per month; Enterprise offers uncapped concurrency.

Vapi (REPORTED). A $0.05 a minute platform (orchestration) fee, with speech-to-text, the LLM, and text-to-speech billed at provider cost, or $0 to Vapi if the customer supplies their own API keys (bring-your-own-key). Ten concurrent lines are included on the Build plan, with additional lines at $10 per line per month. Compliance add-ons were reported at $2,000 a month for HIPAA and $1,000 a month for Zero Data Retention. A commonly cited Deepgram Nova-2/3 streaming STT rate of about $0.0043 a minute is used across multiple sources as the STT line for Vapi-style stacks.

Bland AI (REPORTED, bland.ai/pricing as read July 2026). A bundled per-minute rate covering the LLM, STT and TTS in one number, billed on top of a separate monthly plan fee: Start $0.14/min variable usage, no monthly fee, 10 concurrent calls; Build $0.12/min variable usage plus a $299/month plan fee, 50 concurrent calls, 2,000 calls included; Scale $0.11/min variable usage plus a $499/month plan fee, 100 concurrent calls, 5,000 calls included; Enterprise custom (uncapped concurrency). Telephony is billed separately on every tier, and none of the three tiers' monthly plan fee should be dropped when estimating a real invoice: the per-minute rate alone understates the bill by the full fixed fee. Bland's own pricing materials argue that once provider costs are added, comparable Vapi and Retell stacks land at $0.13–$0.31 and $0.11–$0.25 a minute respectively, which is the vendor's own framing rather than an independent benchmark, though it is directionally consistent with the per-layer figures above.

The four-layer stack

Every voice agent runs the same pipeline regardless of vendor: speech-to-text, an LLM turn, text-to-speech, and a telephony carrier. Formula: all-in cost per minute = platform fee + STT + LLM + TTS + telephony. Telephony (a Twilio-class carrier) was reported at roughly $0.013 a minute for voice minutes, separate from a small monthly phone-number fee.

Three agent shapes

All are ILLUSTRATIVE. A: simple receptionist — short, low-complexity turns, a cheap model, minimal TTS ($0.02/min LLM, $0.02/min TTS). B: support agent with retrieval and tools — longer context, more tokens per turn, a mid-tier model ($0.06/min LLM, $0.04/min TTS). C: outbound sales/qualification — longer calls, a heavier model doing more reasoning per turn, premium voice ($0.10/min LLM, $0.08/min TTS). STT ($0.0043/min) and telephony ($0.013/min) are held constant across shapes.

All-in cost per minute by shape

Bland's bundled rate does not change with model complexity in this comparison because the vendor absorbs that variation into one flat number; Retell's and Vapi's all-in rates rise directly with how much the LLM and TTS layers cost for a given shape. The Bland column above is variable usage only and excludes Scale's separate $499 monthly plan fee, which is fixed regardless of shape and must be added to get a real invoice total (see the next table).

Platform-only versus full-stack cost per 10,000 minutes

The platform-only figure is what a vendor's homepage headline implies; the full-stack figure is what actually lands on the invoice once STT, LLM, TTS and telephony are counted. On shape C, the full-stack cost is 4.6 times (Retell) and 4.9 times (Vapi) the platform-only figure. On Bland, the full invoice at 10,000 minutes ($1,729) is 1.4 times the variable-usage-only figure ($1,230), entirely because of the $499 monthly plan fee, which does not change with shape. Note that Bland Scale's plan includes only 5,000 calls a month; a workload sending 10,000 minutes through many short calls could exceed that limit before it exceeds the minute count, and this article does not have a public overage rate for calls beyond the included allowance (QUOTE-ONLY / UNKNOWN). This also means Scale is not automatically the cheapest or most appropriate Bland tier for every workload: a lower-concurrency, lower-call-count workload might fit Start or Build more cheaply once each tier's own plan fee and call cap are checked against actual usage.

Cost per completed call, and the effect of call duration

Formula: cost per call = all-in cost per minute × average call duration. A short 1-minute call, a mid-length 3-minute call, and a longer 8-minute call, on shape B (support agent):

At 10,000 completed calls a month averaging 3 minutes each (30,000 total minutes), that is $5,169 (Retell), $5,019 (Vapi), or $499 + $3,690 = $4,189 (Bland Scale, plan fee plus variable usage) for the month, before accounting for any failed or abandoned calls that still consume partial minutes. This example itself exceeds Bland Scale's stated 5,000-call monthly allowance (10,000 calls against a 5,000-call plan), so a workload at this volume would need to confirm Bland's overage or Enterprise terms for calls beyond the included count rather than assume Scale's per-minute rate applies unchanged; Retell's and Vapi's per-call economics do not carry an equivalent call-count cap in the sources reviewed.

Reported buyer-side numbers, for cross-check

One independent write-up modeled 40,000 minutes a month across a specific stack choice: Retell (bundled) at roughly $0.07 all-in, about $2,800; Vapi with bring-your-own-key providers (Deepgram Nova-3 STT, GPT-4o-mini-class realtime LLM, ElevenLabs Flash TTS, Twilio) at roughly $0.18–$0.22 all-in, about $7,200–$8,800; Bland Scale at roughly $0.09–$0.11 all-in, about $3,600–$4,400 (REPORTED, one source's specific provider mix, not a universal rate). These figures use a different model and voice mix than the shapes above and illustrate how sensitive the "all-in" number is to which specific STT, LLM, and TTS providers are actually selected.

Sensitivity

1. Model choice. Moving from a cheap model to a frontier model on Retell's own LLM table triples the LLM line (from roughly $0.05–$0.08/min blended to $0.16–$0.32/min for GPT-5.5-class).
2. TTS provider. Retell's own TTS table spans $0.015 to $0.040 a minute, a 2.7x range depending on voice quality tier.
3. Bring-your-own-key versus passthrough. On Vapi, BYOK moves the LLM/STT/TTS spend to the customer's own provider accounts at native rates, avoiding any platform markup on those layers, but does not reduce the $0.05/min platform fee itself.
4. Call length. A workflow that runs longer calls (outbound qualification, complex support) multiplies every per-minute figure linearly; shape and duration compound.

Budgeting traps

• Quoting the platform fee as the whole price. On Retell and Vapi, the platform fee is typically 20–45% of the full-stack cost once the model layers are added.
• Ignoring telephony as a separate carrier bill. All three platforms bill telephony minutes and phone numbers apart from the AI-minute rate.
• Assuming a bundled rate is always cheaper. Bland's flat rate wins on model-heavy shapes (C) but can lose to a cheap-model Vapi or Retell stack on light shapes (A) once BYOK discounts are applied.
• Voicemail and failed calls. A call that reaches voicemail or disconnects early still consumes STT, partial LLM turns, and telephony minutes up to that point.

What to ask before you buy

Ask each vendor for the complete per-minute breakdown by layer, not just the platform fee, and specify your actual model and voice choice. Then multiply by your real average call duration and completion rate, not an assumed round number, before comparing platforms.

ARTICLE 12