A finished 15-second ad for $1.50 in video credits

The expensive mistake in AI video is not a bad take. It is rendering a bad take at delivery resolution, three times, because you were still deciding what the shot was.

I worked the numbers for a fifteen-second spot and the whole thing — twenty directions, three deliverables — came to 150 credits. That is a dollar fifty at the published rate, and most of it was spent on shots that never left the draft tier.

The run

Run the same twenty directions at 1080P and the exploration alone is 300 credits, with another 90 for finals. The exploration is where the money goes, not the delivery.

Why the cheap tier is not a compromise

At 480P you are judging things that do not depend on pixel count:

• Composition — is the subject where the frame wants it.
• Blocking — does the action read in the order you wrote it.
• Timing — does the cut land on the beat.
• Motion — does the camera move do what the prompt said.

None of those improve with resolution. What improves is the finish: edges, texture, the small details a deliverable gets judged on. So this is not "cheap version, then expensive version". It is two different questions asked in the right order.

The shell loop I actually use

# 12 explorations, 5s each, 480P — 60 credits total
for i in $(seq 1 12); do
  curl -s https://api.callirra.com/v1/videos \
    -H "Authorization: Bearer $CALLIRRA_API_KEY" \
    -H "Content-Type: application/json" \
    -d "{\"model\":\"minimax-h3\",\"prompt\":\"$(sed -n "${i}p" shots.txt)\",\"resolution\":\"480P\",\"duration_seconds\":5}" \
    | tee -a jobs.jsonl
done

Then one question per take, written down: does the motion read, yes or no. Takes that fail on motion never get a second render — a longer take of the wrong move is still the wrong move.

Where the plan breaks

You need to judge text or logos. Render one 768P take of the candidate before committing. Readable type at 480P is not a fair test.

The output is vertical. The arithmetic holds, but check framing at the final aspect ratio early; a 16:9 draft of a 9:16 shot is a different composition.

The client wants a look test. That is a different job: one take at 1080P or 960P, not twelve.

The shot is longer than ten seconds. Standard 1080P stops at ten. Plan the finish on a tier that runs to fifteen.

The free path, priced honestly

A new account gets 18 credits in the first week, then 15 a week. At 480P that is 18 seconds in week one and 15 a week after; at 768P, 9 seconds and 7. That is enough to test three five-second directions and finish nothing — which is the correct reading of a free tier. Plan week one as exploration, not delivery.

Why a draft-heavy pipeline is safe

Failed, cancelled and expired jobs release their reservation automatically. Credits go when a file exists. So the failure mode of a bold exploration is "no take", not "a take you paid for and cannot use".

FAQ

Can I upscale a 480P draft instead of re-rendering? You can, and you will be inventing detail. Re-rendering at a higher tier is the cleaner route when the output is the deliverable.

Does 1080P accept every duration? Standard 1080P runs to ten seconds; longer shots need a tier whose ceiling is fifteen.

Is there a cheaper draft tier than 480P? No. One credit per output second is the floor.

Do subscription credits change the arithmetic? The tier rates are identical on every plan; the price of a credit varies with plan and cycle.

The full six-tier rate card is on Callirra's H3 page. If you have a pipeline that beats this arithmetic, the numbers are the interesting part — send them.