Cost model

"You CAN do this, but how much are you willing to spend?"

This doc answers that question with numbers.

The problem

A 5-minute video at 60fps = 18,000 frames, which is 18,000 vision calls if you analyze each one. Most of those frames are near-duplicates: the same scene a fraction of a second apart.

The solution: extract less, resize smaller, cap hard

Frame count comparison (5-minute video)

Mode	Setting	Frames	Vision API calls
Every frame	60fps	18,000	18,000
Interval	1s	300	300
Interval	5s	60	60
Scene (default)	threshold 0.4	~15-25	~15-25
Keyframe	auto	~30-60	~30-60
Scene + cap	threshold 0.4, max 10	10	10

The Every frame and Interval rows are exact math. The Scene and Keyframe rows are content-dependent estimates: a static screen-share produces far fewer frames than a fast-cut montage. With --max-frames, the worst case is fixed regardless of content.

Resize impact

Source	Resolution	JPEG size	Tokens (approx)
No resize	1920x1080	~500KB	high
Resize 512px	512x288	~25KB	low
Resize 256px	256x144	~8KB	minimal

Default resize is 512px. Sufficient for "describe what's happening" prompts. Increase with --resize 1024 for detail work (OCR, small text).

Whisper transcription

Video length	Segments (30s)	Whisper calls
1 min	2	2
5 min	10	10
30 min	60	60
1 hour	120	120

Whisper calls scale linearly at one per 30-second chunk, independent of resolution. They sit well below vision calls per unit cost, so transcription is rarely the line item to optimize.

Example scenarios

Scenario 1: Quick summary of a 5-minute demo video

bun run src/cli.ts analyze video.mp4 \
  --mode scene --scene-threshold 0.4 \
  --max-frames 15 --resize 512 \
  --prompt "Summarize what's shown in this frame"

~15 vision calls (capped)
10 whisper calls
Total: ~25 API calls

Scenario 2: Detailed analysis of a 1-minute clip

bun run src/cli.ts analyze clip.mp4 \
  --mode interval --interval 2 \
  --max-frames 30 --resize 768 \
  --prompt "Describe all text and UI elements visible"

~30 vision calls (capped)
2 whisper calls
Total: ~32 API calls

Scenario 3: Just transcribe a 1-hour lecture

bun run src/cli.ts transcribe lecture.mp4 --segment-duration 30

0 vision calls
120 whisper calls
Total: 120 API calls (all Whisper)

Scenario 4: Smart analysis of a 30-minute meeting recording

bun run src/cli.ts smart meeting.mp4 --max-timestamps 5 --prompt "Describe the diagram or visual shown"

60 whisper calls (30s chunks)
1 text LLM call (timestamp selection)
5 vision calls (targeted frames only)
Total: 66 API calls, 5 of them vision

Blind analyze with scene detection on the same 30-minute recording produces an estimated 75-150 vision calls (content-dependent). Smart mode fixes the vision count at --max-timestamps instead, trading it for one text LLM call. The difference grows with video length, since the vision count stops tracking duration.

Dry-run before model calls

bun run src/cli.ts analyze video.mp4 --dry-run

Prints:

Estimated frame count
Model being used
Resize dimension
Whisper segment count

No API calls made. Use this to check the call count before every run.