By Sagar Shankaran, Founder of CallSphere
Vapi customers report 7 surprise line items beyond the $0.05/min headline. Here is the audit and how CallSphere bundles them away.
Key takeaways
Vapi's $0.05/min platform fee is the first of at least seven recurring line items that show up on a real production deployment. Buyers report unexpected charges for STT, LLM tokens, TTS characters, telephony, number rental, observability, and engineering on-call. CallSphere's flat-tier pricing collapses all seven into a single invoice — no surprise fees, no end-of-month token spikes, no separate observability vendor.
Let's be precise: every line item in this post is documented somewhere in someone's pricing page. None of them are deceptive. They feel hidden only because the headline number that gets shopped — Vapi's $0.05/min — isn't the number that lands on the invoice. The hidden part is the assembly cost.
This post audits the seven recurring line items Vapi customers most commonly report, in rough order of dollar impact.
graph TD
H[Headline: Vapi $0.05/min] --> L1[1. LLM tokens]
H --> L2[2. TTS characters]
H --> L3[3. STT seconds]
H --> L4[4. Telephony minutes]
H --> L5[5. Phone number rental]
H --> L6[6. Observability/logging]
H --> L7[7. Engineering on-call]
L1 --> INV[Real invoice ~$0.33/min + carrying cost]
L2 --> INV
L3 --> INV
L4 --> INV
L5 --> INV
L6 --> INV
L7 --> INV
style H fill:#ffd
style INV fill:#fccFigure 1 — The headline is $0.05. The invoice has seven line items.
The biggest single cost in a voice AI call is usually the LLM. At GPT-4o-realtime rates, the equivalent per-minute cost is about $0.10–$0.18. Verbose system prompts, large RAG contexts, and back-and-forth tool calls all push this up.
Token billing is inherently variable. A confused caller who repeats themselves three times can double the LLM cost of a call. A complex tool-calling chain can triple it. Forecasting this is genuinely hard.
CallSphere bundles the LLM cost inside the flat tier. Verbose conversations are absorbed into the envelope. Token-spike months disappear from finance dashboards.
Every word the agent speaks is metered. ElevenLabs Turbo v2 lists at ~$0.18 per 1,000 characters. A 10-minute call where the agent speaks 7,000 characters costs **$1.26 in TTS alone** — about 25x the Vapi platform fee for the same minute.
Premium voices (the ElevenLabs voices buyers actually want) cost more. Voice cloning costs more. Multilingual voices cost more.
CallSphere bundles premium ElevenLabs voices on appropriate tiers. No per-character meter. No "we'd love to use the better voice but it doubles our spend" conversation.
Deepgram Nova-2 is the de facto STT for voice AI; it lists at roughly $0.0077 per audio minute. Cheaper STTs exist; they tend to perform worse on accents, ambient noise, or vertical jargon (medical terminology, addresses, surnames). Most production deployments end up on Nova-2 or Whisper.
CallSphere bundles STT. Healthcare deployments use vocabularies tuned for clinical terminology; real estate deployments use vocabularies tuned for street names and price formats — without a separate STT bill.
Twilio Programmable Voice prices US inbound at ~$0.014/min, US outbound at ~$0.022/min, plus regional surcharges and toll-free differences. The number itself is a small monthly fee but multiplies across DIDs.
International numbers are dramatically more variable. Vapi's number inventory is heavily US/CA-focused, so international deployments often require BYO Twilio with global numbers — adding complexity.
Hear it before you finish reading
Talk to a live CallSphere AI voice agent in your browser — 60 seconds, no signup.
CallSphere bundles Twilio numbers and inbound/outbound minutes within the tier. Real estate clients in NZ, salons in the UK, and clinics in Canada are supported with regional numbers in the tier price.
Each Twilio DID is roughly $1/month. A multi-location business with 20 numbers (per-clinic, per-salon, per-region) is $240/year just in number rental. Toll-free numbers cost more. Vanity numbers cost dramatically more.
CallSphere bundles standard local numbers at no per-DID add-on. Toll-free and vanity numbers are available on Scale and Enterprise.
Vapi includes basic call recording and a dashboard. Production-grade observability — searchable transcripts, sentiment analysis, lead scoring, intent extraction, escalation flags, RBAC, audit logs, retention controls — is not included. Buyers often add a separate observability tool (Datadog, custom build, or specialized voice-AI eval platforms) for an additional $200–$2,000/month.
CallSphere ships post-call analytics built-in: GPT-4o-mini analyzed sentiment, lead score, intent, satisfaction, escalation flag — surfaced to non-technical operations staff who can grade calls without an engineer. Healthcare ships with 20+ database tables and full audit trails. See /industries/healthcare.
The most expensive line item is the one that doesn't appear on any invoice. With five vendors stitched together, something will degrade at some point. When it does — and a customer is on the line — somebody has to triage in real time.
This is typically 0.1–0.25 FTE of senior engineering carrying cost. At fully-loaded $180k/year, that is $18k–$45k/year.
CallSphere absorbs the on-call. One vendor, one SLA, one escalation path.
| # | Line item | Typical Vapi monthly | CallSphere |
|---|---|---|---|
| 1 | LLM tokens | Highly variable; $1,400 @ 10K min | Bundled in tier |
| 2 | TTS characters | ~$1,200 @ 10K min | Bundled in tier |
| 3 | STT seconds | ~$80 @ 10K min | Bundled in tier |
| 4 | Telephony minutes | ~$200 @ 10K min | Bundled in tier |
| 5 | Number rental | ~$5–$50 | Bundled in tier |
| 6 | Observability | $200–$2,000 | Built-in |
| 7 | Engineering on-call | $1,500–$3,750 | ~0 |
Profile: 8 reception lines across a regional dental group, 8,000 minutes/month.
| Item | Vapi all-in | CallSphere |
|---|---|---|
| Platform | $400 | — |
| LLM tokens | $1,120 | — |
| TTS chars | $960 | — |
| STT | $64 | — |
| Telephony | $160 | — |
| Number rental | $8 | — |
| Observability tool | $400 | Built-in |
| Engineering 0.15 FTE | $2,250 | — |
| Vapi total | ~$5,362 | — |
| CallSphere Growth tier | — | flat ~$1,400 |
| Savings | — | ~$3,962/mo, ~74% |
Beyond raw dollars, the seven-line itemization carries a procurement tax most buyers underestimate. Each vendor needs:
For a 200-person company adding voice AI, the legal and finance overhead of onboarding five vendors can be 30–60 hours of work before a single call rings.
CallSphere's one-invoice model is one security review, one DPA, one renewal, one reconciliation cadence.
graph LR
A[New voice AI project] --> B{Vendor model?}
B -->|Vapi| C[5 security reviews]
B -->|CallSphere| D[1 security review]
C --> E[5 DPAs]
D --> F[1 DPA]
E --> G[5 renewals to track]
F --> H[1 renewal to track]Figure 2 — Procurement overhead, before the first call.
They are documented on each vendor's pricing page, but the total is not surfaced anywhere on Vapi's marketing. Buyers anchor on $0.05/min and discover the seven-line reality post-deployment.
Yes, but free TTS quality is noticeably worse — robotic prosody, mispronounced names, awkward pauses. Most production deployments end up on ElevenLabs or Cartesia, which meter per character.
CallSphere uses best-in-class providers (GPT-4o-realtime, ElevenLabs, Whisper, Twilio) under the hood. The difference is consolidated pricing, single SLA, single observability surface, and zero customer engineering overhead.
Volume discounts exist but require committed spend at each vendor — usually $10K+/month at OpenAI and ElevenLabs to access better rates. CallSphere passes through aggregated volume pricing inside the flat tier.
For production voice AI? Yes. Without searchable transcripts, sentiment, and lead scoring, ops teams can't grade calls without engineering involvement. CallSphere ships this in the box; standalone tools cost $200–$2,000/month.
Vapi assistants can break when upstream vendors change APIs or have outages. With five upstream vendors, the surface area is wide. CallSphere consolidates the surface into one team.
Still reading? Stop comparing — try CallSphere live.
CallSphere ships complete AI voice agents per industry — 14 tools for healthcare, 10 agents for real estate, 4 specialists for salons. See how it actually handles a call before you book a demo.
Beyond the seven invoice line items, regulated buyers (healthcare, financial services, insurance) pay an eighth fee that doesn't appear on any pricing page: the compliance drag of qualifying five vendors instead of one.
In healthcare, every vendor that processes PHI must execute a BAA. Five vendors = five BAAs. Each must be reviewed by InfoSec and legal. Each adds to the sub-processor map your auditors examine.
In financial services, each vendor adds to the third-party risk management surface. audits become more expensive in proportion to the number of in-scope vendors.
CallSphere's Healthcare product ships HIPAA-ready with a single BAA covering the full stack. Compliance teams need one approval, not five. This often shaves 4–8 weeks off go-live timelines for regulated buyers.
Each of the five infrastructure vendors has a "real production" tier that buyers usually don't realize they need until they're live:
These tier climbs can collectively add $1,000–$3,000/month to a 10,000-minute deployment beyond the per-minute meters. They show up after go-live, when you discover the entry-level tier didn't include something you needed.
graph TD
A[Day 1: Free / entry tiers] --> B[Day 30: Hit rate limit]
B --> C[Upgrade OpenAI tier]
C --> D[Day 45: Need custom voice]
D --> E[Upgrade ElevenLabs tier]
E --> F[Day 60: Need SLA]
F --> G[Upgrade Vapi tier]
G --> H[Cumulative tier creep ~$1-3K/mo]
style H fill:#fccFigure 3 — Tier creep across vendors compounds quickly post-launch.
Profile: 12-person commercial insurance brokerage, ~5,500 minutes/month for inbound quote intake and renewal reminders. Compliance with state insurance regulations.
Growth tier flat: typically lands at less than half this number, with built-in audit logging, multi-tenant isolation, and ops-graded transcripts.
It's worth spending a bit more time on what each line item feels like during a real production cycle — because the dollar number alone underestimates the pain.
LLM token spikes happen when long-tail callers trigger long, multi-turn conversations. A confused or angry caller can drive a single call's token cost 4–6x higher than a typical call. At scale, the tail of the distribution drives a disproportionate share of monthly LLM cost. CallSphere absorbs this in the flat tier; Vapi customers either eat the variance or build aggressive token-throttling logic that degrades agent quality.
TTS character spikes correlate with agent verbosity changes. A system prompt update that asks the agent to confirm more details ("Just to confirm, your name is...") can increase TTS character count 20–30%. Engineering teams often discover this only when the ElevenLabs invoice arrives.
Twilio number rental sprawl is silent but real. Multi-location businesses accumulate DIDs over time — campaign-specific numbers, vanity numbers, regional overflow numbers. A 50-clinic operation can easily have 80+ active DIDs at $1+/month each, plus toll-free or vanity at $5–$25/month each.
Observability subscription tier creep happens as call volume grows. Datadog or similar APM tools price by data ingestion. Voice AI generates a lot of data — transcripts, audio metadata, function-calling traces, latency metrics. Subscription tiers climb with usage, often surprising buyers.
Engineering on-call is the line item nobody budgets explicitly until it bites. A weekend incident with a customer on the line, a vendor outage during a campaign launch, a Deepgram model deprecation — each requires a senior engineer to triage. Over a year, this load is genuinely 0.1–0.25 FTE for SMB enterprise.
CallSphere's flat tier makes all five fees disappear from the buyer's mental ledger.
A common pushback to this audit is "fine, those fees exist, but my team will manage them." That's true for the first 30 days. By month six, the cognitive overhead of seven recurring meters becomes background noise — and that's exactly when the fees compound silently. Token usage drifts up after a system prompt change; TTS character count grows as agents handle more complex flows; observability subscription tier climbs as call volume grows.
The fees are hidden not because they are deceptive, but because they are distributed. No single vendor's invoice is alarming. The aggregate is.
CallSphere's flat-tier model removes the cognitive overhead by design. There is one number to track. If usage grows, you size up to the next tier on a planned cadence. There is no slow-creep tax on the AP team.
We will run the seven-line audit on your Vapi deployment and quote a CallSphere tier that beats your current all-in.
Written by
Sagar Shankaran· Founder, CallSphere
LinkedInSagar Shankaran is the founder of CallSphere, where he builds production AI voice and chat agents deployed across healthcare, hospitality, real estate, and home services. He writes about agentic AI, LLM engineering, and shipping voice agents that handle real calls in production.
See how AI voice agents work for your industry. Live demo available -- no signup required.
Oilfield-services vendors in Basra field procurement and tender calls at all hours from operators and international partners. CallSphere answers 24/7 in Arabic and English and routes serious tenders to a human.
A founder's guide to texto a voz (text-to-speech in Spanish): LATAM vs Castilian voices, free options, and how CallSphere ships Spanish agents.
A founder's guide to the Siri voice generator landscape: how AI voice cloning works, what is legal, and how CallSphere uses 57+ voices in production.
A founder's guide to the female voice generator landscape: AI female voices, Japanese voices, robot voices, and how CallSphere ships 57+ voices live.
A founder's guide to AI voice assistants for ecommerce: customer service, order lookup, and how CallSphere fits in versus virtual receptionists.
Phone answering services in 2026 are mostly AI. Here is the real comparison: cost, coverage, languages, and how to pick the right one for your business.
© 2026 CallSphere Inc. All rights reserved.
Made within San Francisco
Watch how CallSphere handles real customer calls, schedules appointments, and processes payments — live.
Try Live DemoBook a DemoCalculate Your ROI