Tavus developer pricing: how are minutes billed (minimums/rounding), and how do I estimate cost for my use case?

Most developers don’t fail on model quality—they fail on economics. You can get a Tavus AI Human to feel human, but you also need to know exactly what that “presence” will cost per user, per session, per month. This guide breaks down how Tavus developer pricing works, how minutes are typically billed (including common minimums and rounding patterns in real-time video systems), and how to estimate cost for your specific use case.

Note: Tavus updates pricing over time. Always double-check the latest details in your Developer Account dashboard or by contacting Tavus sales for exact numbers. This article is meant to help you reason about how minutes accrue and how to model costs, not to quote final rates.

---

The Quick Overview

• What It Is: Usage-based pricing for real-time, face-to-face AI Humans, where you pay for interaction time (minutes) across perception, speech, LLM, TTS, and real-time video rendering.
• Who It Is For: Developers and product teams embedding Tavus into apps via API, building AI SDRs, AI support agents, in-product copilots, onboarding flows, and other live video experiences.
• Core Problem Solved: You get human-level presence—seeing, hearing, and responding in real time—without building and running a massive multimodal stack yourself, and you can predict what that will cost per active user or per conversation.

---

How Developer Billing Typically Works

From a systems perspective, Tavus runs a real-time pipeline—perception → ASR (speech recognition) → LLM → TTS → real-time rendering (Phoenix-4) → video streaming. Pricing is usage-based around that pipeline, usually represented as minutes of AI Human time.

At a high level:

1. You open a real-time session

   • Your app connects a user to an AI Human via WebRTC or a real-time streaming interface.
   • Billing typically starts once the session is established and the Tavus agent is “present” (audio/video live and able to respond).

2. You accrue billable minutes while the AI Human is active

   • Every second that Tavus is rendering the AI Human, listening, and ready to speak counts toward usage.
   • Behind the scenes this covers perception (Raven-1), conversation orchestration (Sparrow-1), and Phoenix-4 rendering.

3. The session ends and usage is rounded

   • When you disconnect or programmatically end the call, total duration is calculated.
   • Most real-time communication platforms bill in time slices (e.g., per second with a small minimum, or per 30/60-second increments). Tavus’ dashboard will show exactly how your minutes are computed and rounded.

Because Tavus is designed for sub-second latency and natural turn-taking, think in continuous conversation time, not just “speaking time.” The AI Human is a live presence, not on-demand clips.

---

How Minutes Are Typically Billed (Minimums & Rounding)

Tavus does not publicly list every internal billing rule, but if you’ve worked with real-time video APIs (Twilio, Agora, Daily), the pattern will feel familiar. Here’s how to reason about it.

1. Session-Based Measuring

• Each connection to an AI Human (one user in a call with one Tavus agent) is a session.
• Sessions have:
  • A start time (when the live agent is available and media is flowing).
  • An end time (when you disconnect or when the server closes the session).

Billable minutes = (end time – start time), adjusted for rounding/minimums.

2. Minimum Billable Duration

Most real-time systems apply a small minimum because spinning up perception + rendering is non-trivial.

Expect something like:

• Per-session minimum: e.g., 30–60 seconds of billing, even if the session ends earlier.
• This protects against dozens of 2–5 second low-value pings, which are expensive to serve.

How to model it:

• If your product uses very short calls (e.g., “tap to ask a 10-second question”), assume each call bills as at least the minimum (e.g., 0.5–1.0 minute).
• If your calls are multi-minute, the minimum barely matters in your cost.

3. Rounding Granularity

Real-time platforms often bill in:

• Per-second increments, sometimes with a 1-second or 10-second minimum;
• Or per-30 or 60-second increments (e.g., 1.1 minutes rounds to 1.5 or 2.0 minutes).

Because Tavus optimizes for “at the speed of human interaction,” per-second accounting is natural for internal systems. Your actual invoice, however, may:

• Show usage summed in minutes with a given rounding rule.
• Aggregate all sessions over a billing period, then round at the plan level.

If you need to be conservative in planning, model round-up to the next 30-second or 1-minute block per session.

4. What Counts as “AI Minutes”?

In typical Tavus-style usage, anything where the AI Human is:

• Visible and/or audible to the user;
• Listening, perceiving, and ready to respond;
• Rendering facial behavior (Phoenix-4) with micro-expressions and natural idle states;
• Running perception on voice, screenshare, or surroundings.

…is part of your billable window.

Non-interaction states like your backend preparing data before a session starts do not count; only real-time agent presence does.

---

Estimating Cost for Your Use Case

You don’t need a final price sheet to model your economics. You need to translate product behavior into minutes per user per month.

Here’s a step-by-step way to do that.

Step 1: Define Your Conversation Pattern

For each key feature in your product, answer:

• Average session length (in minutes)
  Example: a customer success AI Human that runs 12-minute check-ins.
• Sessions per user per month
  Example: 4 check-ins per month → 48 minutes per user/month.
• Number of concurrent users or active users
  Example: 1,000 users actively using this flow.

Formula:
Monthly minutes = average session length × sessions per user per month × active users

Example A — AI SDR / Sales Agent

• 8 minutes per sales call
• 200 calls per month
• Monthly minutes = 8 × 200 = 1,600 minutes

Example B — In-App Onboarding Copilot

• 5-minute sessions
• 3 sessions/user/month
• 500 active users
• Monthly minutes = 5 × 3 × 500 = 7,500 minutes

You can then apply your plan’s per-minute price to get a cost range.

Step 2: Account for Minimums and Rounding

If your calls are short and frequent, minimums matter more.

Assume:

• 1-minute minimum per session, rounded up to the next 30 seconds (conservative planning).

Now compare:

• A 5-minute average session → rounding impact is small.
• A 20-second quick-check interaction → billed as 1 minute.

Example C — Quick FAQ AI in a Product

• Raw average: 0.4 minutes (24 seconds) per session
• Sessions: 3,000 per month
• With 1-minute minimum: billable = 1 × 3,000 = 3,000 minutes
• Without minimum: 0.4 × 3,000 = 1,200 minutes

Your cost model should assume the higher number unless sales/Docs tell you otherwise.

Step 3: Consider Latency and Idle Time

Tavus emphasizes sub-second latency and natural turn-taking, so users will often stay connected for the entire “conversation” (even when neither is actively speaking).

That means:

• Silence still counts if the AI Human is present and listening.
• Long idle states when the user is away should be minimized by:
  • Automatically ending sessions after N seconds of no user audio.
  • Requiring an explicit “start conversation” action before connecting Tavus.
  • Running background LLM tasks off-call when possible.

Design tip: Use system-driven timeouts to cap useless minutes. For example, end the session after 60–90 seconds of no user speech or interaction.

Step 4: Add Headroom for Growth

Once you have a base estimate:

• Add 20–30% margin for:
  • Users who talk longer than average.
  • Spikes from product launches.
  • Internal testing and staging environments.

This gives you a realistic total expected minutes per month.

---

Example Cost Models by Use Case

Below are illustrative examples to help you think in “Tavus minutes.” Plug your own numbers in where needed.

1. AI SDR / Lead Qualification

• 7-minute average conversation
• 150 calls/month per agent
• You simulate 10 human SDRs with AI Humans

Minutes/month = 7 × 150 × 10 = 10,500 minutes

Use-case nuances:

• Sessions are longer, minimums don’t hurt much.
• You’ll care more about conversion per minute than raw cost—if the AI SDR books meetings reliably, per-minute costs are easy to justify.

2. Onboarding or Training AI Human

• 10-minute guided onboarding
• 2 sessions per new user
• 3,000 new users/month

Minutes/month = 10 × 2 × 3,000 = 60,000 minutes

Use-case nuances:

• Usage tied to growth / new signups.
• Easy to cap via product design (fixed flows, scripted duration).

3. Internal Support or IT Helpdesk AI

• 4-minute average troubleshooting session
• 3 sessions per employee per month
• 2,000 employees

Minutes/month = 4 × 3 × 2,000 = 24,000 minutes

Use-case nuances:

• High-volume but manageable.
• You can steer users to text-only flows for simple tickets and reserve AI Humans for complex, multimodal issues (screenshare, tone-sensitive conversations).

---

Features & Benefits Breakdown

Core Feature	What It Does	Primary Benefit
Real-time AI Humans	Streams lifelike video agents with Phoenix-4 rendering in sub-second loops.	Lets you charge by “conversation minutes,” not dev time.
Multimodal Perception	Uses Raven-1 to interpret voice, screenshare, and surroundings.	Each minute is higher value because the agent truly “gets” context.
Flexible APIs	Developer Account exposes APIs and SDKs to embed AI Humans in your app.	You only pay for minutes you actually deliver to users.

---

Ideal Use Cases

• Best for high-touch, repeatable conversations: Because Tavus shines when you need presence—sales calls, support, onboarding—and per-minute pricing maps well to units like “per call” or “per user/month.”
• Best for products where trust and tone matter: Because every billed minute includes perception of micro-expressions, timing, and nonverbal cues—not just text prediction.

---

Limitations & Considerations

• Very short, high-frequency queries can be inefficient: If your typical interaction is a 5–10 second yes/no answer, per-session minimums will inflate your effective per-minute rate. Consider batching micro-questions into a single “session” or using text-only flows for trivial queries.
• Predictability depends on your UX: If users can idle in conversations indefinitely, your minutes will spike. Use timeouts, explicit “end call” buttons, and clear conversational boundaries to keep usage within your budget.

---

Pricing & Plans (Developer Accounts)

Tavus offers Developer Accounts specifically for teams building with APIs and real-time video agents. While exact dollar figures and tiers may change, the structure typically looks like:

• Usage-based billing on minutes of AI Human interaction
• Optional tiers or committed-use plans with:
  • Volume discounts at higher minute thresholds.
  • Enterprise options for “over 2 billion interactions” scale.
  • SLAs with enterprise uptime guarantees.

For the latest and exact pricing:

• Check the billing section in your Tavus Developer dashboard.
• Or contact Tavus for a tailored quote if you’re planning high-volume or enterprise deployment.

A typical pattern you’ll see is something like:

• Builder / Pay-as-you-go: Best for early-stage developers testing ideas, pilots, and low-volume apps. You pay only for minutes used, with minimal commitment.
• Scale / Enterprise: Best for teams rolling out AI Humans to production across sales, support, or product workflows, needing volume pricing, advanced SLAs, and dedicated support.

---

Frequently Asked Questions

How exactly are Tavus minutes calculated for developers?

Short Answer: Minutes are calculated based on the duration of active real-time sessions where an AI Human is present, listening, and able to respond, typically measured per session with a small minimum and rounding.

Details:
Every time your app connects a user to an AI Human, a session begins. The clock starts when the AI Human is live (audio/video active, perception running) and stops when the session ends. All real-time workload—perception, speech recognition, LLM, TTS, and Phoenix-4 rendering—is captured in this duration.

Most real-time systems apply:

• A per-session minimum (to avoid abusive micro-sessions).
• Rounding to a defined granularity (e.g., per second, or up to the nearest 30/60 seconds).

Your Tavus dashboard will expose your usage in minutes; your invoice reflects those minutes multiplied by your plan’s per-minute rate. For mission-critical budgeting, verify the exact rounding rule with Tavus support or sales.

---

How do I estimate Tavus cost for my specific use case?

Short Answer: Break your product into conversation types, estimate minutes per session, sessions per user, and active users, then apply a conservative rounding assumption and add 20–30% headroom.

Details:
Follow this pattern:

1. Identify each conversation type (sales call, onboarding, support, etc.).
2. For each type, estimate:
   • Average session length (in minutes).
   • Sessions per user per month.
   • Number of active users or total sessions per month.
3. Compute raw minutes = length × sessions × users.
4. Adjust for minimums and rounding:
   • If sessions are short (<1 minute), model each as at least 1 minute.
5. Sum across all conversation types.
6. Add 20–30% buffer to cover longer-than-expected conversations and traffic spikes.

Once you have that total minute estimate, multiply by the per-minute rate from your Developer Account plan or Tavus quote. This will give you a workable monthly cost model to compare against your revenue, LTV, or internal savings.

---

Summary

Building with Tavus isn’t about rendering a face on a chatbot—it’s about running full-stack human computing at the speed of real conversation. Developer pricing reflects that: you pay for minutes of real-time presence, not static content.

If you can describe:

• How long your users talk to AI Humans,
• How often they do it,
• And how many users you expect,

you can model your Tavus cost with surprising precision. From there, it’s a simple question: does each minute of lifelike, multimodal interaction generate more value than it costs? For most high-trust, high-touch workflows—sales, support, onboarding—the answer is yes.

---

Next Step

Get Started