What $0.05/min actually covers
The $0.05/min Vapi fee pays for the orchestration layer: the real-time pipeline, endpointing, barge-in, call routing, dashboards, transcripts, recordings, and observability. It does not include STT, LLM, TTS, or the phone-number carrier — those are pass-through, meaning Vapi charges you what the underlying provider charges (with managed) or you bring your own key (BYO). The orchestration fee is the smallest line item on almost every call, but it's the one that makes the rest of the pipeline usable in production.
The other line items
Here's the rough 2026 breakdown for common provider choices. Prices change frequently — always sanity-check against each provider's current rate card.
- STT (Deepgram Nova streaming) — roughly $0.0043/min at 2026 pricing.
- STT (Whisper via OpenAI) — roughly $0.006/min but higher latency.
- LLM (GPT-4o mini streaming) — call it $0.03–$0.05/min on typical agent turns.
- LLM (GPT-4o full streaming) — closer to $0.15–$0.25/min on chatty flows.
- TTS (ElevenLabs Turbo streaming) — roughly $0.05–$0.10/min depending on words spoken.
- TTS (Deepgram Aura) — roughly $0.03/min, lower latency, fewer voice options.
- Telephony (Twilio US local) — $0.014/min inbound + outbound leg.
Worked example: a typical 3-minute support call
Take a 3-minute inbound support call using GPT-4o mini, Deepgram Nova, ElevenLabs Turbo, and Twilio US: 3 min × ($0.05 Vapi + $0.005 STT + $0.04 LLM + $0.07 TTS + $0.014 telephony) ≈ $0.54 all-in. Push to GPT-4o full and ElevenLabs Multilingual v2 and the same call is closer to $1.20. Downgrade to a smaller LLM + Deepgram Aura TTS and you're at $0.30. The 4x range across identical duration is entirely provider selection.
Outbound campaigns and concurrency
Outbound is where costs multiply. A 5,000-call campaign averaging 90 seconds per contact at $0.40/call is $2,000 in variable cost. Add the marketing team's headcount, the CRM sync, and the transfer-to-human minutes, and the true per-lead cost lands around $2–$5. This is why outbound teams obsess over first-word latency (shorter turns → shorter calls) and prompt tightness (fewer LLM tokens → cheaper turns).
Where teams overspend
The two biggest cost leaks are running the biggest LLM on every turn (use a smaller model on the hot path, reserve the frontier model for tool-call decisions) and leaving TTS non-streaming (you pay the same but audio starts later, so callers ask you to repeat, doubling turn count and duration). The third is verbose system prompts — every extra 1,000 tokens costs real money at scale because it's retokenized on every turn.
How to model your own workload
Track three numbers per call: turns, average words per response, and total call duration. Multiply against provider rates. Vapi's dashboard exports these so you can back-solve unit economics before scaling a campaign. Build a simple spreadsheet: rows are provider combos, columns are duration buckets, cell is expected cost. Update it monthly as providers cut rates — 2026 has already seen two rounds of LLM price cuts.
BYO vs managed provider pricing
Managed adds a small markup (usually 5–15%) but you skip all the credentials and rate-limit configuration. BYO gives you direct provider pricing and any volume discount you've negotiated. The crossover point is usually around 20k minutes/month; below that, managed pays for itself in engineering time. Above it, BYO saves enough to justify the ops overhead.
Hidden costs to budget for
Beyond the per-minute rate, budget for: number rental (Twilio is ~$1/mo per US number), recording storage (small but grows), webhook egress if your tools live on a different cloud, and human transfer time (a warm transfer keeps the meter running on both legs). None of these are individually large; combined, they're often 10–15% of the raw per-minute cost.
Comparing Vapi's fee to raw provider math
It's worth stress-testing the $0.05/min orchestration fee against a hypothetical DIY build. To replicate Vapi on raw providers you'd need: a WebSocket bridge to Twilio (engineer-weeks), streaming glue between Deepgram and OpenAI (engineer-weeks), an endpointing state machine (engineer-weeks), barge-in with cancellable TTS (engineer-weeks), tool-call plumbing (engineer-weeks), a dashboard for transcripts and latency (engineer-months), and on-call for provider incidents. At a fully-loaded engineer cost of ~$250k/year, six engineer-weeks is $30k. That pays for 600,000 minutes of Vapi orchestration. Almost no team should build this in-house before hitting that threshold — and by then you know exactly what to build.
Cost per completed goal, not cost per minute
Per-minute cost is a means, not the metric that matters. What matters is cost per completed goal: cost per qualified lead, cost per booked appointment, cost per resolved support ticket. A cheaper agent that resolves 40% of tickets is more expensive than a costlier agent that resolves 75%, because unresolved tickets escalate to human agents at $2–$5 per interaction. When you evaluate a stack change, compute the completion-rate delta before the per-minute delta.
Budgeting for growth
Voice agent cost curves are non-linear because concurrency limits, provider volume tiers, and telephony rate bands all step-function at scale. A pilot that costs $500/month can quickly become $50,000/month if a marketing team switches on a broad outbound push. Set spend alerts on the Vapi dashboard, on each provider's dashboard, and on Twilio. Treat the first month of production as a discovery exercise: watch which flows dominate cost and iterate on those specifically.
Worked example: 1,000 calls at scale
Suppose you run 1,000 support calls a day averaging 2.5 minutes each — a modest inbound operation. That's 2,500 minutes daily, ~75,000 minutes a month. On a mid-tier stack (Vapi + Deepgram Nova + GPT-4o mini + ElevenLabs Turbo + Twilio US), per-minute cost lands around $0.18 all-in: $0.05 Vapi + $0.005 STT + $0.04 LLM + $0.07 TTS + $0.014 telephony + ~$0.001 overhead. Monthly variable cost is 75,000 × $0.18 ≈ $13,500. Swap in GPT-4o full for accuracy on tool-call turns only (say 30% of turns) and the LLM line rises to ~$0.07 blended — total climbs to ~$16,000. Downshift TTS to Deepgram Aura and total drops to ~$11,000. On the same 1,000-call day, that's a $5,000 monthly swing on TTS alone. Now stress-test outbound at the same 1,000/day but averaging 90 seconds and requiring GPT-4o full for qualification reasoning: 45,000 minutes at $0.28/min ≈ $12,600/month. Add a 20% warm-transfer rate to a human, averaging 45 extra seconds of overlap per transfer, and you tack on ~$500. Model your specific workload before signing an outbound contract; the difference between a $10k and a $30k month is entirely provider mix.
Vapi vs Retell vs Bland per-minute rate comparison
Headline orchestration fees look different across the three platforms, and the shape of the bill matters as much as the sticker rate. At typical 2026 rates: Vapi charges $0.05/min plus pass-through provider costs, so a mid-tier stack lands around $0.18/min all-in. Retell charges roughly $0.07/min at its default tier plus pass-through, so the same mid-tier stack lands around $0.20/min — a small premium for a more curated experience. Bland's flat rate is roughly $0.09/min bundled at low volume, dropping to $0.05–$0.06/min at enterprise commit tiers — and because the stack is bundled you don't stack provider markups, so a heavy campaign that would cost $0.30/min all-in on Vapi with GPT-4o full can be closer to $0.09/min on Bland. Break-even usually lands around 50k minutes/month: below that, Vapi's pass-through model is cheapest; above it and especially at outbound-heavy patterns, Bland's flat pricing wins. Retell sits in the middle and rarely wins on price alone — its argument is qualitative (out-of-the-box naturalness) rather than economic. When you run the math, remember to include number rental, storage, and hidden overhead — those add roughly 10–15% on any platform.