If you sell a product that generates avatar videos, every video your customers make costs you money. A seat-based SaaS can mostly ignore marginal cost; a video product cannot. Founders who skip this math usually discover it when a few heavy users push a plan below zero margin.
This post builds a unit economics model for an AI avatar product from scratch. All customer numbers are hypothetical and exist only to show the arithmetic. The provider costs use PitchTwin's public pack prices, so you can swap in your own volumes and plans.
The three numbers that matter
Unit economics for an avatar product comes down to three inputs:
- Cost per second of output video. What your provider charges you.
- Seconds of video per customer per month. How much your customers actually generate, including re-renders.
- Price per customer per month. What you charge.
Everything else, such as payment fees, free trials and support, sits on top of these.
Step 1: cost per second of video
PitchTwin sells prepaid tokens that never expire. One token is one second of video when you bring your own audio, and 1.5 tokens per second when PitchTwin voices the script in the avatar's cloned voice.
| Pack | Price | Tokens | Cost per token |
|---|---|---|---|
| Starter | $9 | 600 | $0.0150 |
| Growth | $39 | 3,000 | $0.0130 |
| Scale | $149 | 12,000 | $0.01242 |
From there, cost per video is simple multiplication. For a 45-second video:
Own audio: 45 s × 1.0 token/s = 45 tokens
Our voice: 45 s × 1.5 token/s = 67.5 tokens
At Scale ($149 ÷ 12,000 ≈ $0.01242 per token):
Own audio: 45 × $0.01242 ≈ $0.56 per video
Our voice: 67.5 × $0.01242 ≈ $0.84 per video
If you use audio mode with your own text-to-speech provider, add that provider's cost per second to the own-audio figure. Do not compare the two modes on token price alone.
Step 2: a hypothetical plan
Assume you run an AI SDR tool. Your mid-tier plan looks like this:
- Price: $99 per month
- Allowance: 50 personalized videos per month
- Average video length: 45 seconds
- Voice: PitchTwin voices the script
Worst case: every customer uses the full allowance
Tokens per customer: 50 videos × 67.5 tokens = 3,375 tokens
Video cost: 3,375 × $0.01242 ≈ $41.92
Payment processing also comes off the top. At Stripe's standard US card rate of 2.9% + 30¢ per successful charge:
Payment fee: $99 × 0.029 + $0.30 ≈ $3.17
Gross margin at full usage:
$99.00 − $41.92 − $3.17 = $53.91
$53.91 ÷ $99.00 ≈ 54.5% gross margin
That is workable, but it is well below the 75 to 80 percent many SaaS founders are used to planning around.
Realistic case: average usage below the cap
Most customers do not use their whole allowance every month. Suppose your own usage data shows the average customer generates 60% of the allowance, so 30 videos:
Tokens: 30 × 67.5 = 2,025 tokens
Cost: 2,025 × $0.01242 ≈ $25.15
Margin: $99.00 − $25.15 − $3.17 = $70.68
$70.68 ÷ $99.00 ≈ 71.4%
The gap between 54.5% and 71.4% is the whole reason allowances exist. Model both cases, and price so the worst case is still positive.
Switching to your own audio
If you generate audio elsewhere and send it in audio mode, the token cost at full usage drops:
Tokens: 50 × 45 = 2,250 tokens
Cost: 2,250 × $0.01242 ≈ $27.95
Now add your text-to-speech provider's cost for 37.5 minutes of audio. Whether this is cheaper overall depends on that provider's pricing and plan. For how audio mode works, see using your own ElevenLabs audio with an avatar API.
Step 3: the costs people forget
Re-renders
Users edit a script and render again. Those re-renders are real usage. If your logs show that 1 in 10 videos is a re-render, your effective allowance is 55 renders, not 50:
55 × 67.5 tokens × $0.01242 ≈ $46.11 at full usage
Failed generations are different: PitchTwin refunds them automatically, so platform failures do not add to your cost.
Free trials
Suppose each trial user gets 3 videos of 30 seconds, voiced:
3 × 30 s × 1.5 = 135 tokens
135 × $0.01242 ≈ $1.68 per trial user
If 5% of trial users convert to paid, the trial video cost per paying customer is:
$1.68 ÷ 0.05 = $33.60
That is about a third of the first month's revenue on the $99 plan. Treat it as part of acquisition cost and decide whether 3 videos is the right trial size.
Pack size
Buying in larger packs lowers cost per token. On the same full-usage plan:
Starter: 3,375 × $0.0150 = $50.63
Growth: 3,375 × $0.0130 = $43.88
Scale: 3,375 × $0.01242 ≈ $41.92
Since tokens never expire, buying the largest pack you can reasonably use carries no risk of losing unused balance.
Sensitivity: how provider price changes the picture
The provider's per-minute price moves margin more than anything else. Here is the same full-usage plan (37.5 video minutes per customer) at several per-minute rates. Competitor figures are public list prices as of October 2026; check the provider's site, and note that features and included voice differ by provider.
| Provider rate | Source | Cost for 37.5 min | Gross margin on $99 (after $3.17 fee) |
|---|---|---|---|
| $1.12/min | PitchTwin Scale, our voice | $41.92 (exact token math) | 54.5% |
| $2.97/min | Synthesia Creator, approx. (source) | $111.38 | negative |
| $4.00/min | HeyGen Avatar IV (source) | $150.00 | negative |
Those providers have real strengths, including mature studio tools and large avatar libraries, and they can be the right choice for low-volume or studio-led products. For a plan that bundles dozens of personalized videos per customer, though, the per-minute rate decides whether the plan works at all. A fuller comparison is in AI avatar API pricing compared.
Track real cost per customer
Assumptions are a starting point. Once you have traffic, record the exact cost of every video. Each finished generation returns tokens_charged and duration_ms:
import os
import requests
API = "https://pitchtwin.ai/api/v1"
HEADERS = {"Authorization": f"Bearer {os.environ['PITCHTWIN_API_KEY']}"}
COST_PER_TOKEN = 149 / 12000 # Scale pack
def record_cost(generation_id: str, customer_id: str, db) -> None:
gen = requests.get(f"{API}/generations/{generation_id}", headers=HEADERS, timeout=30).json()
if gen["status"] != "succeeded":
return
db.insert(
"video_costs",
customer_id=customer_id,
generation_id=gen["id"],
seconds=gen["duration_ms"] / 1000,
tokens=gen["tokens_charged"],
cost_usd=round(gen["tokens_charged"] * COST_PER_TOKEN, 4),
)
Run this from your webhook handler after verifying the signature (see the webhooks docs). A monthly query over video_costs grouped by plan tells you your real gross margin per plan, and which customers sit above the line.
Checklist
- Compute cost per video for each mode you offer.
- Price each plan so full-allowance usage still leaves a positive margin.
- Model average usage from data once you have it.
- Count re-renders and trial videos as cost.
- Include payment fees.
- Log
tokens_chargedper customer from day one.
For turning these numbers into plan design, read how to price an AI video feature in your SaaS.
Try it
Sign up for 60 free seconds of video with no card, render a test video, and check tokens_charged on the result to confirm the math against your own scripts. The API docs cover every field used above.