$0.26–$0.31 to audit an hour of video

Cheap first, expensive where it matters

Each dot is one way of reading the boards. The big chart sets large boards found against cost: up is better, right is more expensive. Below it, the same for small, distant boards, and time against cost, where lower left is cheaper and faster. Cost and time are on logarithmic scales, so each gridline step multiplies rather than adds.

A two-pass plan. Run Apple Vision on every hour, then send only the high-priority hours to Gemini Flash-Lite, which found 85% of boards against Apple's 69%.

What each option costs at 100,000 hours

Every option reads one frame per second, which is 360 million frames. Prices are list prices with no volume deal. Accuracy is the share of sponsor boards found on the 156 frames we labelled by hand.

Three things stand out:

• Apple's reader is the cost winner. At 1 frame per second it needs about 25,167 Mac host-hours, $26k to $31k.
• Video Intelligence is the most expensive option we measured. On its own it costs $7.20 to $9.00 per video hour and found 45% of boards.
• Gemini finds the most boards, including 61% of the small ones, at $148k to $295k. That is 4.8–9.6× the Apple price. Whether the extra boards are worth it depends on the client.

The whole pipeline, stage by stage

One hour of video, from the video file to per-brand seconds, on one Mac host with 8 workers. Each range covers three timed runs on an otherwise idle Mac: full-match hour, 1080p30 (14.2 min per hour); 2022 final highlights, 1080p50 (15.1 min per hour); 2026 group highlights, 720p60 (12.5 min per hour).

Not in these numbers:

• Human review of detections (precision audit, recall labels): not timed, not costed
• Moving the video into AWS and storing it: not priced
• A 24-hour minimum applies to every EC2 Mac host, so a host must stay busy for the per-hour figures to hold

How fast

One Mac host processes an hour of video in about 12 to 15 minutes with 8 workers, or 21 to 24 minutes with 4. The table shows calendar time for 100,000 hours on more hosts, at the slower speed of each.

Does it hold up on full matches?

The cost model was built on highlight clips, so we ran two complete matches, 30 hand-labelled frames each. All frames were live play. These samples are small, so the ranges are wide.

Of 150 detections checked by eye per match, 131 were real in the Brazil match (fused, 0.87) and 144 in the Portugal match (Apple only, 0.96). Video Intelligence on its own was weaker on the full match, with 11 of 20 of its single-method detections real.

How this compares with what is on the market

Prices are per video hour at one frame per second, from each vendor's public page, checked on 2026-10-15. Sponsorship vendors do not publish prices, so only the general-purpose APIs can be compared on cost. Nobody publishes a numeric accuracy figure, so ours is the only recall and precision number in the table.

Why ours is cheaper and faster

1. We rent a computer, not a service. Cloud APIs add margin, support and scaling to every call.
2. Apple's OCR runs on the Mac itself. No upload, no queue, no network.
3. The job is narrow. We read text and match a short list of brands, and Apple's OCR does that well enough.
4. It runs as a batch on a dedicated machine, with most cores busy and nobody else in line.
5. We skip what vendors sell: valuation, reports, live delivery and human review.

The trade is accuracy: we find 69% of boards, Gemini 85%. We are batch, not live.