Comparison

Agni vs Vapi: Real Costs After Stacking API Keys

Vapi's $0.05/min looks attractive until you add Deepgram for STT, OpenAI for the LLM, ElevenLabs for voice, and Twilio for telephony. Here's what the real bill looks like.

RM
Rahul MehtaVP Engineering, Ravan.ai
7 June 2025  ·  7 min read
Agni vs Vapi: Real Costs After Stacking API Keys

Vapi is one of the most popular developer tools in the voice AI space globally. Its appeal is real: an elegant API, support for swapping any STT/LLM/TTS provider, and a base infrastructure fee that looks attractively low. But for Indian businesses evaluating total cost of ownership, the stacked billing model has a way of producing a very different number from the headline rate.

Let's build the real Vapi bill for a typical Indian outbound deployment.

How Vapi's Billing Works

Vapi is an orchestration layer. It connects your STT provider, your LLM provider, your TTS provider, and your telephony provider — and charges for the orchestration itself. The component providers bill you separately. This means you need:

  1. A Vapi account (platform fee)
  2. An STT provider API key (e.g., Deepgram, AssemblyAI)
  3. An LLM API key (e.g., OpenAI GPT-4o, Anthropic Claude)
  4. A TTS/voice provider API key (e.g., ElevenLabs, PlayHT)
  5. A telephony provider (e.g., Twilio, Vonage) — separate SIP configuration

The advantage is flexibility: you can mix and match providers. The disadvantage is that you're managing 4–5 vendor relationships, 4–5 billing accounts, and 4–5 failure points — and the total cost is the sum of all of them.

Building the Real Vapi Bill for India

Let's use a realistic Indian outbound deployment: 10,000 calls per month, 3 minutes average duration = 30,000 minutes per month.

Vapi Platform Fee

Vapi charges approximately $0.05/min for its orchestration layer (at the time of writing). Plus a monthly platform fee depending on plan. Estimate: $0.05 × 30,000 = $1,500/month in Vapi platform costs alone.

STT: Deepgram Nova-2

Deepgram Nova-2 at $0.0043/min: $0.0043 × 30,000 = $129/month. Note: Deepgram has limited training on Indian accent data, which affects accuracy for Hinglish and regional Indian English. You may need Nova-2 General, not the cheaper streaming tier.

LLM: OpenAI GPT-4o

GPT-4o pricing varies by token count. A typical voice AI call generates 500–1,500 tokens. At $5/1M input tokens and $15/1M output tokens, estimate $0.01–0.02 per call. For 10,000 calls: $100–200/month.

TTS: ElevenLabs

ElevenLabs professional tier: approximately $0.18 per 1,000 characters. A 3-minute call generates roughly 600–900 characters of TTS output. Per call: ~$0.14. For 10,000 calls: ~$1,400/month.

Telephony: Twilio (India outbound)

Twilio's outbound call rate to Indian mobiles: approximately $0.035–0.05/min. For 30,000 minutes: $1,050–1,500/month.

Total Vapi stack for 10,000 calls/month at 3 min average:
Vapi orchestration: ~$1,500
Deepgram STT: ~$129
OpenAI LLM: ~$150
ElevenLabs TTS: ~$1,400
Twilio telephony: ~$1,275
Total: ~$4,454/month ≈ ₹3.74 lakh/month
Effective per-minute rate: ~₹12.5/min

The Agni Bill for the Same Volume

Agni Scale plan: ₹9,999/month platform fee + ₹4.5/min, all-in.

For 10,000 calls at 3 minutes average: 30,000 minutes × ₹4.5 + ₹9,999 = ₹1,44,999/month.

Effective per-minute rate including platform fee: ₹4.83/min.

Monthly savings vs Vapi stack: ₹3.74 lakh – ₹1.45 lakh = ₹2.29 lakh/month. Annual: ₹27.5 lakh.

The Hidden Costs the Numbers Don't Show

Beyond direct cost, the Vapi stack creates operational overhead that has real cost:

  • Engineering time: Integrating 4–5 providers, managing SIP configuration, debugging cross-provider issues — estimate 1–2 engineer-days per month in ongoing maintenance
  • Compliance gaps: None of the Vapi stack components are India-hosted by default. Deepgram processes in US. OpenAI processes in US. ElevenLabs processes in US. Each creates a DPDP cross-border transfer violation for Indian customer data.
  • Hinglish accuracy: No component in the standard Vapi stack is trained specifically on Indian code-switching. STT accuracy for Hinglish on Deepgram is measurably lower than Agni's purpose-trained Indian language models.
  • Vendor risk: 5 separate vendors means 5 separate points of outage, API change, or pricing change

When Vapi Makes Sense

Vapi's model is excellent for developers who need maximum control — running their own fine-tuned LLM, testing novel voice combinations, building a research prototype, or deploying in a market where the specific component mix matters. If you're a developer who wants to own every layer, Vapi is well-designed for that.

For an Indian business that wants to deploy outbound calls in Hinglish, needs DPDP compliance, and wants a predictable monthly bill, the flexibility tax is simply too high.

Ready to get started?

Get a single, transparent bill for your Indian voice AI. Start at app.ravan.ai or write to us at info@ravan.ai.

ComparisonVapiPricingAPI CostsVoice AIIndia

Ready to deploy voice AI that speaks India?

Agni handles Hinglish, regional dialects, RBI-compliant call flows, and sub-300ms latency — built specifically for Indian enterprises.