Pay Once. Ask Forever.
Content creator, student, rubric scorer, security team, moderation screener: whoever you are, a media library only earns its place if the application uploads it, stores it, understands it, and lets you keep asking about it without a new bill each time. Securely, accurately, as often as you need, with no token math and no extra charge just to ask a question. We checked how the big AI models handle each of those steps, and here's why Echosaw stands out.
Ask most people whether AI can "watch" a three-hour recording and they'll say yes. A few models can. Gemini takes about three hours at its default resolution. TwelveLabs' Pegasus takes two. That settles the first thing anyone building a media library asks: can I upload it?
It doesn't settle the rest. Cataloguing footage, collecting a semester of lectures, scoring calls against a rubric, going through camera recordings and screening uploads before they go live each bring four more considerations:
- Storage: where does the video live after I upload it?
- Distillation: what do I actually get back from it?
- Cost: what do I pay each time I ask about it?
- Recall: can I find it again next month, alongside everything else I've recorded?
The matrix
| Echosaw | Google Gemini API | TwelveLabs Pegasus 1.6 | Amazon Nova Lite / Pro | NotebookLM | OpenAI GPT API | Anthropic Claude API | |
|---|---|---|---|---|---|---|---|
| Upload | |||||||
| Hours of video in one file | ✓ 3.5 hrs, 5 GB | ✓ ~3 hrs at low res | ✓ 2 hrs, 10 GB | ◐ 1 GB cap | ◐ 200 MB cap | ✗ no video input | ✗ no video input |
| Storage | |||||||
| Keeps your video and its analysis | ✓ until you delete it | ◐ deleted after 48 hrs, unless you host it in your own bucket | ✓ within plan storage | ◐ you host it in your own S3 | ✓ in the notebook | n/a | n/a |
| Distillation | |||||||
| Watches the picture, not just the audio | ✓ | ◐ 1 frame per second; Google warns fast motion can be missed | ✓ | ◐ 960 frames total: one every ~11 sec on a 3-hr video | ✗ audio transcript only | n/a | n/a |
| Structured results: timeline, timestamped transcript, detections | ✓ timeline, transcript, entities, people, objects, signals, location | ✗ free-form text | ◐ text + timed segments | ✗ free-form text | ✗ transcript text | n/a | n/a |
| Moderation with timestamped evidence and publish gating | ✓ built in | build it yourself | build it yourself | build it yourself | ✗ | n/a | n/a |
| Score a recording against your own rubric or policy | ✓ built in | build it yourself | build it yourself | build it yourself | ✗ | n/a | n/a |
| Study tools from your own recordings | ✓ study packets, spaced review, classes with quizzes, tests and roster mastery | build it yourself | build it yourself | build it yourself | ◐ flashcards and quizzes | n/a | n/a |
| Cost | |||||||
| Ask as often as you like after analysis | ✓ questions aren't metered | ✗ every question is billed (~$0.81 for a 3-hr video) | ◐ monthly usage allowance | ✗ every question is billed | ◐ free tier with usage limits | n/a | n/a |
| Recall | |||||||
| Search your whole library, answers linked to the moment | ✓ right moment first: 13 of 14 | ✗ one request at a time | ◐ index search; right moment first: 4 of 14 | ✗ one request at a time | ◐ one notebook at a time | n/a | n/a |
✓ yes · ◐ partly · ✗ no · "build it yourself" = a general model you can prompt, with no built-in workflow · n/a = the model can't take video. Moment results are from our published bench (TwelveLabs searched an index holding only the test video). NotebookLM wasn't on the bench. Sources at the bottom of the post.
Upload: the easy part, for a few
Two of the best-known model APIs, OpenAI's and Anthropic's, don't take video at all. You extract frames yourself and send pictures. Of the models that do take video, Amazon Nova caps a file at 1 GB and NotebookLM caps an upload at 200 MB. Gemini and Pegasus take long files. Echosaw takes up to 3.5 hours in a single file.
Storage: where the video goes afterward
A file uploaded through Gemini's Files API is deleted after 48 hours. To keep it, you register it from your own Google Cloud Storage bucket and run that bucket yourself. Nova reads from your S3 bucket the same way. Either way, the model doesn't hold your library. You do.
Echosaw keeps every file and its analysis until you delete it. There's no expiry clock.
Distillation: what the model actually saw
Long video is where models start skipping. Past 16 minutes, Amazon Nova samples 960 frames, total. On a three-hour recording that's one frame about every eleven seconds. The ten seconds in between are never looked at. Gemini samples one frame a second, and Google's own documentation warns that fast motion and quick scene changes can be missed. NotebookLM accepts the video but keeps only a transcript of the audio, so anything shown and never said out loud is gone.
Even when a model does see the video, you get text back: a summary, an answer, timestamps if you ask. Echosaw analyzes the audio and the picture once and keeps the results as structured data you can browse: an intelligent timeline, a timestamped transcript, detected entities, people, objects and signals, and location data. Moderation flags come with timestamped evidence, and flagged media stays private. Recordings can be scored against your own rubric or policy, and turned into study packets with spaced review.
Recall: asking again next month
The model APIs answer about the file in front of them. They keep no library, and you can't ask "which of my 40 recordings covered the budget?"
Echosaw searches every recording in your account and links each answer to the moment it came from. On our published video-search bench, Echosaw returned the right moment first for 13 of 14 questions in a 17-item library. TwelveLabs, searching an index that held only the test video, returned it first for 4 of 14. The harness, questions and raw responses are public.
More up front, less over time
Model APIs charge you every time you ask. Gemini bills about 100 tokens per second of video at its default resolution, so every question about a three-hour video sends about 1.08 million tokens. On Gemini 3.8 Flash that's around $0.81 a question today, and Google's list price doubles on January 1. Its file is gone after 48 hours, so next week you upload and pay again. Keeping the video cached so it's ready to answer costs about $389 a month for that one recording.
Echosaw charges once, when the video is analyzed. A 3-hour video on the Agency plan costs $77.40. After that, questions to Emma aren't metered. Share it through an organization, where there are no per-seat fees, so that $77.40 covers every question your whole team asks about the video. The plan covers your library, storage, search, Emma, the study and teaching tools, and API and MCP access. The more you use a recording, the more that one charge is worth.
| Number of questions asked about one 3-hour video | Echosaw (Agency) | Gemini 3.8 Flash, today | Gemini 3.8 Flash, from Jan 1 |
|---|---|---|---|
| 10 | $77.40 | $8.10 | $16.20 |
| 50 | $77.40 | $40.50 | $81.00 |
| 100 | $77.40 | $81.00 | $162.00 |
| 500 | $77.40 | $405.00 | $810.00 |
Echosaw: one-time analysis at $0.43/min audio+video, plus the $49/month Agency plan. Gemini: input tokens only, static mode at default resolution, before any caching discount, with no re-upload costs counted. Output tokens are extra. And that Gemini spend still buys no timeline, no detections and no searchable library.
The short version
- Upload: most AI models can't take hours of video at all, and the ones that can cap the length or the file size. Echosaw takes up to 3.5 hours in a single file.
- Storage: elsewhere, uploads expire or sit in a storage bucket you run yourself. Echosaw keeps every file and its analysis until you delete it.
- Distillation: elsewhere, long video gets sampled, seconds get skipped or only the audio is kept, and you get plain text back. Echosaw analyzes the video and the audio into a timeline, a timestamped transcript and detections, with moderation, rubric scoring and study tools built in.
- Cost: elsewhere, every question is billed. Echosaw charges once, when the video is analyzed, and your questions aren't metered.
- Recall: elsewhere, answers cover one file at a time. Echosaw searches your whole library and links every answer to its moment.
Sources: Gemini video understanding, Files API, caching, pricing · TwelveLabs Pegasus 1.6, pricing · Amazon Nova video understanding, Bedrock price list (us-east-1) · NotebookLM sources, studio outputs · OpenAI and Anthropic API documentation · Echosaw pricing and FAQ.