B-Roll API benchmarks
How often each service gives the right shot first. The same 100 requests for every service, two judges from two companies, and every result public.
Subjects and scenes
People, companies and products, places, events, archival footage and generic scenes: 85 requests.
| Service | Right shot first (Gemini / Claude) | Right shot in top 5 | Median time | Price per 1,000 searches | Rights |
|---|---|---|---|---|---|
| Clypmint, web mode | 70.6% / 56.5% | 78.3% | 7 s | $120 | Marked not cleared |
| Brave image search | 44.7% / 44.7% | 74.7% | 0.7 s | $5 | No license data |
| Clypmint, licensed | 47.1% / 41.2% | 49.4% | 5 s | $20 | Licensed, with credit line |
| Pexels | 21.2% / 24.7% | 33.5% | 0.4 s | Free | Stock license |
| Wikimedia Commons search | 10.6% / 9.4% | 11.8% | 0.4 s | Free | Licensed |
| Openverse | 7.1% / 7.1% | 8.8% | 1.9 s | Free | Licensed |
Movie and TV scenes
One named scene from a film or a show, the clips creators post: 15 requests. A moment inside a video is judged on its own frame, cut the way your machine would cut it.
| Service | Right shot first (Gemini / Claude) | Right shot in top 5 | Median time | Price per 1,000 searches | Rights |
|---|---|---|---|---|---|
| Clypmint, web mode | 60% / 53.3% | 66.7% | 7 s | $120 | Marked not cleared |
| Brave image search | 53.3% / 46.7% | 73.4% | 0.6 s | $5 | No license data |
| Clypmint, licensed | 0% / 0% | 0% | 4.3 s | $20 | Licensed, with credit line |
| Wikimedia Commons search | 0% / 0% | 0% | 0.4 s | Free | Licensed |
| Openverse | 0% / 0% | 0% | 1.5 s | Free | Licensed |
| Pexels | 0% / 0% | 0% | 0.4 s | Free | Stock license |
By category
Right shot first, average of the two judges.
| Category | Clypmint, web mode | Clypmint, licensed | Wikimedia Commons search | Brave image search | Openverse | Pexels |
|---|---|---|---|---|---|---|
| People (15) | 53.4% | 26.7% | 3.4% | 46.7% | 0% | 0% |
| Companies and products (15) | 53.4% | 23.4% | 0% | 30% | 3.4% | 30% |
| Places (15) | 76.7% | 70% | 6.7% | 60% | 26.7% | 63.4% |
| Events (15) | 76.7% | 56.7% | 23.4% | 50% | 0% | 0% |
| Archival footage (10) | 70% | 50% | 30% | 65% | 10% | 0% |
| Generic scenes (15) | 53.4% | 40% | 3.4% | 23.3% | 3.4% | 36.7% |
| Movie and TV scenes (15) | 56.7% | 0% | 0% | 50% | 0% | 0% |
How we ran it
The requests
100 requests in seven categories. A model (Gemini 3.5 Flash) wrote them from a fixed list of categories before any service was run, the way a video editor types them. The list was frozen: no request was changed or removed after results came in. Three of them name a person or a scene a little wrong, and they stayed in. Every service was scored on them the same way. The prompt and the list are in the downloads.
The services
Each service got the request as written and returned its top 5. Clypmint ran through its public engine, once with licensed media only and once in web mode. When nothing it found reached its own bar, Clypmint returned its closest checked shots, which are free, and those were judged like any other result. Wikimedia Commons ran through its search API, Openverse and Brave through their image search, and Pexels through its photo and video search, one of each in turn.
The judges
Two models judged every result, blind to the service: Gemini 3.5 Flash (Google) and Claude Sonnet 5.5 (Anthropic). A result counts only if it shows exactly what was asked and is a real photo or film frame, not a drawing, a render or a stand-in. Each judge saw what a customer gets: a photo's own picture, a video's thumbnail, for a moment inside a video the frame at its middle, and for a live website a frame of the recording our command makes of it. The two judges agreed on 88.2% of first results. Clypmint's own engine checks shots with a Gemini model, so the Claude column is the one with no stake in it.
Price and speed
Price is the list price per 1,000 searches (Clypmint's credits are prepaid, from $10). Time is the median from sending the request to the answer, from one machine in the US.
When
All services were run and judged on October 25, 2026.
Check it yourself
- queries.json: the 100 requests and what a right shot shows. queries-prompt.json: how they were written.
- results.json: every service's top 5 for every request, with both judges' verdict on each.
- harness.mjs, judge2.mjs and summarize.mjs: the code that ran the services, judged them and made this page.
What it does not show
- We made this test and we sell one of the services. That is why the requests, results and code are public.
- 100 requests is a small sample, and Clypmint's answers vary a little from run to run: a second run moved its scores by up to 5 points. Read a gap under 5 points as a tie.
- Judges are models, and they miss things. Each verdict is in the download, so you can check any of them.
- The free services are faster. A Clypmint search takes seconds because a model checks every shot before you get it.
- Web results are not cleared for reuse. Their rights belong to the owner.
Try the same requests
Try them on /broll without a key. Credits start at $10.