How it works
Numbers
Pricing
Log in
Ask video anything.
VideoSense watches, reasons about what it sees, and answers in plain language — with the clip and the chart to prove it. Below is a real session, replayed.
VideoSense · skydive library Which clips show only freefall — no parachute out? Which clips show only freefall — no parachute out? Working… sql_query · phase-tag scan (12 clips) 4ms analyze_video · #7 → retry (frame gap) 1.9s analyze_video · confirm no-canopy ×3 5.1s show_video · 3 clips to side-channel 0.3s 3 of 12 clips are pure freefall — no canopy in frame.1 Solo freefall over the lake ▸ 0:02 2 Wingsuiter tracking away ▸ 0:04 3 Wingsuit close pass ▸ 0:03 #1 Solo freefall #2 Tracking away #3 Close pass ▸ Steps 12 · sql ×2 · watch ×3 · 8.4s A recorded session — the question is typed, the agent streams its steps, then answers with the matching clips.
500+ videos in the production library, growing daily
2,200+ facts extracted from them, queryable
4,300+ semantic index rows — transcripts, captions, facts
92% eval pass rate across 143 graded tasks
Why it feels different
Just ask No SQL, no dashboards, no tagging. Plain-language questions over your whole library.
It actually watches A multimodal model reads the video content — not just filenames or metadata.
Answers you can see The conclusion first, then the playable clips and the chart that back it — with its steps one click away.
The answer, and what it cost VideoSense · skydive library Which clips show only freefall — no parachute out? 3 of 12 clips are pure freefall — no canopy in frame.#1 ▸ 0:02 Solo freefall over the lake #2 ▸ 0:04 Wingsuiter tracking away #3 ▸ 0:03 Wingsuit close pass FREEFALL-ONLY VS CANOPY · ALL 12 CLIPS pure freefall 3 canopy in frame 9 ▸ Steps 12 · sql ×2 · watch ×3 · 8.4s $0.0535 · 92k tok · 9% ctx The anatomy of every answer — the conclusion first, the playable proof, then the receipt.
Measured, not promised
131/143 graded tasks passed — real Gemini, reproducible test world
46/47 must-pass tasks held
6.1s median answer across 143 eval tasks (529 runs)
$0.077 a heavy 4-video query, probed end to end
Pass rate by capability
all 12 scored dimensions · deterministic verifiers, no LLM judge · single run, July 2026
Timestamps 100%
Multi-turn memory 100%
Safety 100%
Identity 100%
State tracking 100%
Tool use 96%
No overreach 95%
Honesty 94%
Counting 92%
No ID leaks 89%
Retrieval 85%
Entity matching 80%
The receipt, itemized
what a real query actually costs
Watch one video ≈60k tok ≈$0.018
One reasoning turn Gemini flash ≈$0.034
Heavy 4-video query 142k tok · 84s $0.077
And the wait
the long tail is queries that watch several videos end to end
median 6.1s
p90 25.5s
p95 68s
Every number above is measured, July 2026 — the library from production, the scores from the repo's own eval suite. No projections.
Get started
See it on your own library.
Book a 20-minute walkthrough, or get instant access and point it at your videos. Priced per workspace — cancel anytime.
Secure checkout via Stripe · your videos stay private to your workspace.