How our daily videos are made

We may earn a commission if you sign up through a Higgsfield link on this page. It doesn’t change what we write. How this works

Written 13 November 2026. This is the pipeline as it runs today, what it costs from our own spend log (29 Oct to 12 Nov, scaled to 30 days), and what we tried that did not work. Vendor prices are on each tool’s page, linked below, with the day we read them.

The pipeline, in order

1. Script. The daily email is written first. A second step turns it into the episode script, and Claude writes the scripts, translations and cards (about ₹280 a month).
2. Voice. English audio comes from ElevenLabs on its Creator plan; Hindi audio comes from Sarvam. Our own server makes the audio files.
3. Video. Python builders on the same server draw the English and Hindi videos: cards, charts, captions and sound over the voice. Captions are timed by Whisper. Remotion draws the host frames in our tests; the host is not in the daily run yet.
4. Publish. The finished files go to storage, then a message with the links goes to our Slack and Telegram. Posting to YouTube is a separate step that we control.

What it costs a month

From our spend log for 29 Oct to 12 Nov, multiplied by 30/14. Prices of the plans themselves are on the tool pages; this is what we actually used.

If we add a face vendor, our estimate is ₹265 to ₹1,320 more a month, which would put the total near ₹3,700 to ₹4,800 plus the server. That is a plan, not a bill: it is not approved.

The AI video tools, and what we did with each

Higgsfield and, where noted, other links on this page are the ones we use or looked at. Only the Higgsfield link pays us; the rest go to the vendor with no tracking code.

What failed

• A cheaper Higgsfield model did not follow our audio. We tried Seedance Mini first. It made its own speech instead of our recording (the sound matched our line at 0.26 on a 0 to 1 scale, against 1.00 for the model we kept). It is not a lip-sync tool.
• Open models. Hallo takes English audio only. SadTalker is built for real photos and handles our drawn hosts badly. Both ship parts whose licences bar commercial use unless swapped.
• Synthesia. Its photo-avatar rules refuse illustrated characters, which is what our hosts are.
• Our mouth-sync check is a weak judge. Several clips scored close to their wrong-audio control, so we look at frames and listen to them as well.
• Our own face builder crashed twice on the Higgsfield clips (no still moment to cut on). It is not fixed; the host is not in the run until it is.
• Cost does not scale. At about $6.75 a minute on Higgsfield credits, a daily host would cost far more than the ₹1,320 ceiling above. That is why we bake a small set of clips once and reuse them instead of rendering each day.

What we would tell someone starting

Start with the voice and the script, because that is where the viewer decides. Add a face only when a test shows it helps. Our test of two Shorts a day for seven days has not run yet; the pass mark is a median of 100 views in 48 hours, where today it is 8.

Higgsfield pricing and plans → · Higgsfield vs Runway →