Three months ago, a client asked me to spec out an AI voice agent for their dental practice. They wanted to automate appointment confirmations, handle after-hours calls, and route emergencies. Simple enough, right?
I pulled up four vendor pricing pages. Each one quoted a per-minute rate that looked reasonable. Then I started doing the actual math -- telephony costs, LLM token costs, voice synthesis costs, phone number fees, platform fees, overage charges. The "$0.07 per minute" headline price turned into $0.18 per minute fully loaded.
That experience made me angry enough to write this post. AI voice agent pricing is deliberately confusing. Every vendor advertises a low per-minute rate and buries the rest. I'm going to lay out what things actually cost so you can budget properly before signing anything.
The Anatomy of a Voice Agent Bill
Before we compare vendors, you need to understand the five cost layers. Every single AI voice agent involves all five, whether the vendor bundles them or bills them separately.
1. Voice synthesis (text-to-speech). This is the AI generating speech from text. Costs vary based on voice quality -- the more natural-sounding, the more expensive. Basic voices run $0.01-0.02/minute. Premium cloned voices hit $0.04-0.08/minute.
2. Speech recognition (speech-to-text). Converting caller speech into text for the LLM to process. Usually $0.005-0.01/minute: Deepgram's flagship Nova-3 streams at $0.0048/min (its new Flux streaming model is $0.0065/min English). Deepgram and AssemblyAI are the main providers here.
3. LLM processing. The AI brain -- GPT-5.6, Claude, or similar -- figuring out what to say. This is token-based. A typical 3-minute call uses 1,500-3,000 tokens. At GPT-5.6-luna rates ($0.20/$1.20 per million tokens), that's well under a cent per call. At GPT-5.6-terra rates ($2/$12 per million), it's $0.01-0.03 per call. Small individually, adds up at volume.
4. Telephony. Actual phone call costs. Inbound and outbound rates vary by country. US calls: $0.01-0.03/minute. Plus phone number rental: $1-3/month per number. Some vendors include this, many don't.
5. Platform fee. The vendor's software, infrastructure, and margin. This is either baked into the per-minute rate or charged as a monthly subscription. Sometimes both.
Vendor-by-Vendor Pricing Breakdown
Bland.ai
Bland is the most straightforward pricing in the space, which is why we use them for a lot of client work. As of August 2026 their model is plan-tiered, with one genuinely all-inclusive per-minute rate per plan -- LLM, TTS, STT, and telephony are bundled, with no token charges.
- Start: $0/month platform fee, $0.14/minute connected time. Capped at 100 calls/day.
- Build: $299/month, $0.12/minute.
- Scale: $499/month, $0.11/minute.
- Enterprise: Custom rates and commitments.
What they don't tell you: Very little is hidden now -- the all-inclusive rate really does include the model. The nuance is the platform-fee math: the $0.02/min saving on Build only recoups its $299/month fee at roughly 15,000 minutes/month, so most businesses should stay on the free Start plan far longer than the tier names suggest. If you see Bland quoted at "$0.09/minute" anywhere, that's the old pricing -- it's gone.
Real-world cost: For a dental practice running appointment confirmations on the Start plan -- $0.14/minute flat. A 2-minute call costs about $0.28, with nothing added on top.
Synthflow
Synthflow targets non-technical users with a visual builder -- but as of August 2026 it is no longer a self-serve option at all. The company has withdrawn its published subscription tiers; synthflow.ai/pricing now lists a single figure: enterprise contracts starting at $30,000 annually (roughly $2,500/month before negotiation).
- Current pricing: Enterprise-only, from $30,000/year. No published per-minute rate, no self-serve signup.
- Gone: The old $29-$1,400/month tiers and $0.08-$0.15/minute rates that most comparison articles still quote. They are no longer offered to new customers.
What that means for you: if you were considering Synthflow for its no-code builder and premium ElevenLabs voices, the entry price now excludes small businesses entirely. When the self-serve plans existed, we measured real-world costs of $0.28-0.35 per minute of total call duration -- already the most expensive platform per minute. Today the comparison isn't about per-minute rates at all: it's a five-figure annual contract against platforms you can start on for tens of dollars.
Synthflow's white-label features and sub-accounts still make sense for agencies reselling voice AI at enterprise scale. Everyone else should look at Bland, Vapi, or ElevenLabs below.
Vapi
Vapi is the developer-first platform. Think of it as the "build your own" option with maximum control. Their pricing reflects this -- it's modular and transparent, but you need to add up all the pieces.
- Platform fee: $0.05/minute. This is just Vapi's infrastructure.
- Voice (TTS): Bring your own. ElevenLabs Flash/Turbo $0.05 per 1,000 characters (~$0.05 per minute of generated speech), Deepgram ~$0.02/min. (PlayHT, formerly the budget option here, shut down after its team's July-2025 acquisition by Meta.)
- Transcription (STT): Deepgram Nova-3 streaming $0.0048/min, AssemblyAI $0.01/min.
- LLM: Bring your own API key -- Vapi passes model usage through at cost. GPT-5.6-luna ~$0.001/min, GPT-5.6-terra ~$0.02/min, Claude Sonnet 5 ~$0.015/min.
- Telephony: Twilio integration. US local calls $0.0085/min inbound and $0.014/min outbound, plus ~$1/month per number.
What they don't tell you: Nothing, honestly. Vapi is the most transparent pricing in the space. The catch is that you need developer skills to set it up, and you're managing multiple vendor relationships (Vapi + voice provider + LLM provider + telephony provider). If anything breaks, you're debugging across four systems.
Real-world cost: Vapi + ElevenLabs Flash/Turbo + GPT-5.6-luna + Deepgram + Twilio = approximately $0.12-0.14/minute. A 2-minute call costs about $0.24-0.28. With GPT-5.6-terra instead of luna: $0.15-0.18/minute.
ElevenLabs Conversational AI
ElevenLabs entered the voice agent space in late 2025. They're primarily a voice synthesis company, so the voice quality is best-in-class. But their agent platform is newer and less battle-tested.
- Free: a small monthly bundle of agent minutes. Testing only.
- Paid plans (Starter $6/mo → Creator $22 → Pro $99 → Scale $299 → Business $990, monthly billing): each bundles a monthly allocation of Agents minutes that grows with the tier.
- Past your bundle: additional agent minutes bill at $0.08/min (per the November-2025 reshuffle), with LLM token costs passed through separately.
- Enterprise: Custom, with volume discounts.
What they don't tell you: These rates cover voice synthesis only. Telephony is extra (they use Twilio under the hood). LLM costs are separate. And their agent platform, while improving rapidly, doesn't have the workflow complexity of Bland.ai or Vapi yet. Limited branching logic, basic tool calling.
The voice quality, though -- it's unmatched. If your use case demands the most natural-sounding voice (luxury brands, healthcare where empathy matters, high-touch sales), ElevenLabs is where you start. For a full comparison, see our ElevenLabs vs Synthflow breakdown.
Real-world cost: inside your plan's bundled minutes, an agent call costs only the LLM passthrough plus telephony; past the bundle it's $0.08/min + LLM + Twilio. Fully loaded, expect roughly $0.10-0.15/minute once you're paying overage -- a 2-minute call around $0.20-0.30.
Total Cost Examples: 100, 1,000, and 10,000 Calls/Month
Enough theory. Here's what you'd actually pay, assuming an average call duration of 2.5 minutes and a standard budget-tier LLM stack (GPT-5.6-luna class).
100 Calls Per Month (Small Business)
Total minutes: ~250/month. This is a small dental office, a solo real estate agent, or a local service business handling after-hours calls.
- Bland.ai (Start): 250 min x $0.14 all-inclusive = $35/month
- Vapi: ~$33/month (250 min x $0.13) + $1 phone number = $34/month
- ElevenLabs: Creator $22/month (its 275-min Agents bundle covers all 250 min) + LLM passthrough + telephony = ~$30/month
- Synthflow: no longer available at this tier -- enterprise contracts start at $30,000/year
At this volume, the differences are small. You're looking at $30-35/month across the platforms that will actually sell to you. Compare that to a part-time receptionist at $15-20/hour -- even 2 hours per week is $130-160/month. The AI voice agent pays for itself in the first week.
1,000 Calls Per Month (Growing Business)
Total minutes: ~2,500/month. A mid-size medical practice, a regional insurance agency, or an e-commerce company handling order status calls.
- Bland.ai (Start): 2,500 min x $0.14 all-inclusive = $350/month (the Build plan's $299 fee + $0.12/min would total $599 -- stay on Start at this volume)
- Vapi: ~$325/month (2,500 min x $0.13) + $1 phone number = $326/month
- ElevenLabs: Pro $99/month + $101 Agents overage (1,262 min past the 1,238-min bundle x $0.08) + LLM passthrough = ~$220/month; or Scale at $299 with the whole volume inside its bundle
- Synthflow: enterprise-only -- the $30,000/year contract floor works out to ~$2,500/month minimum
Now the spread matters. ElevenLabs' bundled Agents minutes make it the value play at this volume. Bland.ai and Vapi are close behind. Synthflow's enterprise floor is 7-11x the alternatives. At 1,000 calls/month, you're replacing 1-2 full-time phone staff ($3,500-7,000/month). Even the most expensive self-serve option here is over 90% cheaper than human agents.
10,000 Calls Per Month (Enterprise/Call Center)
Total minutes: ~25,000/month. A regional call center, a national service company, or a large healthcare system.
- Bland.ai: published Scale plan puts a ceiling on the math -- $499 + 25,000 min x $0.11 = ~$3,250/month; enterprise contracts at this volume negotiate below that.
- Vapi: ~$2,750/month (25,000 min x $0.11 with volume optimization) = ~$2,750/month
- ElevenLabs: Business $990 (12,375-min bundle) + ~$1,010 Agents overage (12,625 min x $0.08) + LLM passthrough = ~$2,100-2,400/month, or Enterprise custom.
- Synthflow (Enterprise): contracts start at $30,000/year; expect a quote well above the ~$2,500/month floor at this call volume.
At enterprise volume, everything becomes a negotiation. ElevenLabs' bundle-plus-overage math is the best published starting point; Bland and Synthflow both move to custom contracts; Vapi stays competitive but the modular cost structure means you're optimizing across four vendors to keep costs down.
The Hidden Costs Nobody Mentions
The per-minute rates are just the beginning. Here's what catches people off guard:
Setup and configuration: $500-5,000. Building a production-quality voice agent takes 20-80 hours depending on complexity. Prompt engineering, conversation flow design, edge case handling, testing. You're either paying a developer or an agency (like ours) to do this right.
Voice cloning: $0-500. If you want a custom voice that matches your brand, most platforms charge for voice cloning. ElevenLabs charges based on plan tier (free on Scale+). Bland.ai charges a one-time fee. Budget $200-500 for a professional-quality custom voice.
Phone numbers: $1-15/month each. Standard US numbers run $1-3/month. Toll-free: $3-5/month. Local numbers in multiple area codes: $1-2 each. A business with 5 local numbers is paying $5-10/month just for phone lines.
Twilio/telephony markup: 10-40%. Platforms that use Twilio under the hood mark up the telephony costs. Vapi is transparent about this. Others bake it into their rate. If a vendor won't tell you their telephony provider, assume a 30% markup.
LLM cost spikes. Complex conversations use more tokens. A 5-minute call where the customer asks 8 questions and the AI needs to look up data can use 5,000+ tokens. At GPT-5.6-terra rates ($2/$12 per million tokens), that's $0.02-0.05 in LLM costs alone. Multiply by 10,000 calls and you've got $200-500/month in LLM costs that weren't in the vendor quote.
Monitoring and maintenance: $200-500/month. Someone needs to review call logs, update prompts, handle edge cases, and fix issues. This is ongoing. The agent doesn't maintain itself. We charge clients $300/month for basic monitoring and prompt optimization. Doing it in-house takes 5-10 hours/month of someone's time.
How to Cut Costs by 30-50%
We've optimized voice agent costs for dozens of deployments. Here's what actually moves the needle:
Use a budget-tier model for 80% of calls. Most inbound calls are routine -- appointment confirmations, business hours, order status. GPT-5.6-luna or Claude Haiku 4.5 handles these perfectly at roughly 90% lower LLM cost than a frontier model. Route complex calls to GPT-5.6-terra or transfer to a human. This alone cuts LLM costs by 70-80%.
Optimize conversation length. A well-designed conversation flow completes the same task in 1.5 minutes instead of 3. That's 50% savings on every per-minute cost. Get to the point. Confirm quickly. Don't let the AI ramble.
Negotiate annual commitments. Every vendor offers 15-30% discounts for annual contracts. At 1,000+ calls/month, this is worth negotiating. We've gotten 40% discounts for clients committing to 12-month terms with volume minimums.
Use webhooks instead of polling. Some platforms charge for "keep-alive" time while waiting for webhook responses. Structure your integrations to minimize wait states. Every second the call is open costs money.
Batch similar calls. If you're making outbound calls (appointment reminders, follow-ups), batch them during off-peak hours. Some telephony providers offer lower rates for off-peak. It's a small savings per call but adds up at 10,000+ calls/month.
Build vs. Buy: The Real Decision
You have two paths:
Use a platform (Bland.ai, ElevenLabs): Faster to deploy (days, not weeks). Higher per-minute costs. Less customization. Best for businesses under 5,000 calls/month. (Synthflow used to sit in this bucket; it's enterprise-only now.)
Build on Vapi or raw APIs: More setup time (weeks to months). Lower per-minute costs at scale. Full customization. Best for businesses over 5,000 calls/month or with unique requirements.
The break-even point where building your own becomes cheaper than a platform is typically around 3,000-5,000 calls per month. Below that, the development and maintenance costs don't justify the per-minute savings.
There's a third path: hire an agency to build and manage it. That's what most of our AI voice agent clients do. We handle the architecture, integration, and ongoing optimization. They get enterprise-grade voice AI without the technical overhead. Setup runs $2,000-8,000 depending on complexity, plus $300-800/month for management.
The Bottom Line
AI voice agents in 2026 cost between $0.10-0.25 per minute fully loaded, depending on voice quality, LLM choice, and platform. For a small business taking 100 calls/month, budget $30-40/month. For a mid-size company at 1,000 calls/month, budget $220-350/month. For enterprise at 10,000 calls/month, budget $2,100-3,300/month before negotiated discounts.
These numbers are 85-95% cheaper than human agents for equivalent call volumes. The ROI is almost always positive within the first month. The question isn't whether to deploy voice AI -- it's which vendor and architecture give you the best cost-to-quality ratio for your specific use case.
If you want us to run the numbers for your business, reach out. We'll model the total cost across platforms and tell you exactly what to expect. No sales pitch -- just math.