Pricing for AI video is genuinely hard to compare, and not by accident. Providers sell different units — characters, seconds, credits, AI minutes, monthly allowances — and a headline price tells you almost nothing until you convert it into the only unit that matters, which is a finished video you would actually post.
What follows is the arithmetic with our own numbers, including which of them we trust and why.
How each thing is billed
| Step | Typically billed as | Comparable across providers? |
|---|---|---|
| Text to speech | Per character, or a monthly allowance | Only after conversion |
| Voice cloning | One-off, or included | Usually |
| Images or B-roll | Per image | Yes |
| Video generation | Per second of output | Yes, and it dominates |
| Lip-sync or avatar | Per second, or flat per clip | No — two different technologies |
| Captions and assembly | Usually free | Not a cost factor |
The text-to-speech numbers
This is the step everyone assumes is expensive and it generally is not.
Sarvam's bulbul:v3 is ₹30 per 10,000 characters. We list that as a verified figure because we reconciled it against our own Sarvam usage dashboard rather than reading it from documentation — a distinction worth making, since rates published in docs and rates on an invoice have disagreed before.
Cartesia sells an allowance instead: the Pro plan is $5 a month for approximately 133 minutes of audio, taken from their own pricing page on 29 August 2026.
What we can and cannot verify
It is worth being explicit about this, because a confident number in a blog post is exactly the kind of thing that gets repeated and then quoted back at people.
| Figure | Source | Confidence |
|---|---|---|
| Sarvam bulbul:v3, ₹30 / 10K chars | Reconciled against our usage dashboard | Verified |
| Cartesia Pro, $5/mo ≈ 133 min | cartesia.ai/pricing, 29 Aug 2026 | Read from the vendor, not invoiced |
| Image generation, ~$0.003/image | Provider documentation | Estimate — not reconciled |
| Per-second video generation | Provider documentation | Estimate — varies by resolution |
Where the money actually goes
The intuition that voice is the expensive part is wrong, and it leads people to optimise the cheapest step.
Voice for a short reel is a rupee or two. Images are a few rupees. What changes the number by an order of magnitude is anything billed per second of generated video — and above all any avatar or lip-sync step, which is priced per second of output and turns a daily posting habit into a real monthly bill very quickly.
This is why a plan that includes a talking avatar in every video is a fundamentally different cost structure from one that does not, regardless of what the monthly headline says.
How to compare a tool honestly
- Find out what unit it sells — characters, seconds, credits or AI minutes.
- Work out how many finished videos one month of that unit produces at your typical length.
- Divide the monthly price by that number. This is the only figure worth comparing.
- Check whether regeneration is charged. A tool where a second attempt costs again is dearer than its rate suggests, because first attempts are not always usable.
- Check whether a per-second step — avatar, lip-sync, generated video — is included or extra. That is what dominates the total.
Our own numbers, for reference
Bolo is ₹399 a month for 15 videos and ₹899 for 30, which works out to roughly ₹27 to ₹30 per finished video including voice, visuals and captions. There is a free tier for a first video.
We publish this partly because it is useful for comparison and partly because the exercise above is only honest if we apply it to ourselves. See pricing for what each plan includes.
The short version
- Compare per usable video, never per month. The units are deliberately different.
- Sarvam TTS is ₹30 per 10,000 characters — verified against our own dashboard.
- Cartesia Pro is $5/month for about 133 minutes, per their pricing page on 29 Aug 2026.
- A 30-second Gujarati script is 400 to 600 characters, so voice is a rupee or two.
- Per-second video and avatar steps dominate everything else. Check whether they are included.
- Ask whether regeneration is charged. First attempts are not always usable.