Why it matters
Why it matters
A model that goes back to a frame and looks closer is new. Our test shows what it finds on a real job and what that costs per clip.
Why I tried it
Why I tried it
Elorian says Glimpse can reason in visual space, revisiting and refining internal visual representations as thinking unfolds. Andrew Dai's launch post shows it bringing crops, selected frames and new visual views back into its reasoning.
We check every rough cut of our reels for production defects before it goes out. A model that goes back to a moment and looks closer should be good at that, so we gave it that job.
We covered the launch as a News. This is the follow-up: our own test.
How it runs
How it runs
Same clips, same prompt, one call per clip, for every model. Free ffmpeg checks set the floor.
The set.
15 clips from our own rough cuts, 9:16, 20 to 36 seconds each, with 20 defects we planted: 14 you see (7 kinds, such as a frozen avatar, a black frame or a caption over the eyes) and 6 you hear (dropouts, pops, captions out of sync). 3 clips are clean. A find counts when it names the defect within 1 second of where it is.
Round 1, through OpenRouter.
Elorian Glimpse, Gemini 3.8 Flash and Qwen 3.8 Omni Flash, each clip sent as video, provider defaults, no system prompt. On OpenRouter, Glimpse is served by Together, not by Elorian.
Round 2, on Elorian's own API.
The same test with only the route changed: each clip uploaded to Elorian's API, 5 frames a second, default reasoning, free tier. 14 of the 15 clips: we stopped before clip 15 to stay under our cost cap. Clip 15 holds one sound defect only, so every visual defect and all 3 clean clips are in.
The floor.
ffmpeg's own detectors (black frames, freezes, scene cuts, silence, pops), free, on the same clips and scored the same way.
Andrew Dai, launch post on X, 16 Oct 2026Elorian, Introducing Glimpse, 16 Oct 2026Elorian API docsOpenRouter docs, video inputOpenRouter model list and pricesOur News on the launch
What happened
On the 14 clips both rounds ran.
- Visual
- Glimpse on Elorian's API found 8 of 14 visual defects. Through OpenRouter 6, free ffmpeg 5, Gemini 2, Qwen 1.
- Short ones
- Defects under half a second: Glimpse on Elorian's API 3 of 6, ffmpeg 3, Glimpse through OpenRouter 2, Gemini 2, Qwen 0.
- Looked again
- On Elorian's API it named 9 times in 4 clips that sit between its 5 fps samples, such as 7.56 s, the last clean frame before a mouth glitch. Through OpenRouter: 0 of 86 times at that level.
- Timing
- Off by 0.08 s on average on Elorian's API, 0.62 s through OpenRouter, 0.4 s for Gemini, 7.0 s for Qwen.
- Sound
- Glimpse found 0 of 5 sound defects in both rounds. It gets a transcript of the words, not the sound. ffmpeg found 2, Qwen 1, Gemini 0.
- False alarms
- On the 3 clean clips: 3 for Glimpse on Elorian's API, 5 through OpenRouter, 5 for Qwen, 2 each for Gemini and ffmpeg.
- Time
- Per clip: about 446 s (7.4 minutes) on Elorian's API, 68 s through OpenRouter, 28 s for Gemini, 158 s for Qwen, 7 s for ffmpeg.
- Cost
- Per clip: USD 0.503 on Elorian's API (USD 0.018 per second of video), USD 0.074 through OpenRouter, USD 0.014 for Gemini, USD 0.007 for Qwen, nothing for ffmpeg. Round 2 cost USD 7.04 for 14 clips.
- Per token
- USD 0.60 in and 3.00 out per million tokens, 20 % under Gemini 3.8 Flash at 0.75 and 3.75, as Elorian says. It used 22,445 input tokens per second of video, calling the model again over 8 to 31 steps per clip. Gemini used 105. So per clip it came to about 36 times Gemini and 70 times Qwen.
The catches
The catches
Through OpenRouter it does not look again.
The same model through OpenRouter named 0 of 86 times at frame level and used 2,058 input tokens per second of video, against 22,445 on Elorian's API.
The fix If looking again is what you pay for, call it on Elorian's own API.
No video inline.
Elorian's API refuses video sent inside the request. Each clip has to be uploaded as a file first, and deleted afterwards.
The fix Upload with the files endpoint, point the call at the file, delete it when the answer is in.
It hears nothing.
It gets the words as a transcript, not the sound: it said it heard no audio on 14 of 14 clips. Its one sound finding was on a clip with no sync defect.
You cannot see or cap the thinking.
The API returns no reasoning, only step titles, and the reasoning cannot be capped, so the cost of a clip is known only afterwards. Ours ran from USD 0.26 to 0.93.
Skip it.
Our verdict: skip it, unless you have a dedicated use case that needs it to look again at video frames while it reasons.
What this does not tell you
One job, ours: 15 short clips of our own rough cuts (14 in round 2), one call per clip, default settings. A longer video, another kind of defect or other settings can come out differently.
Prices, models and Elorian's free tier as of 31 Oct 2026.
Tested 31 Oct 2026, verdict given 1 Nov 2026. Models and prices change, so this page is dated.
Results may vary
Your own verdict
Four questions decide it for your case.
Does your job need it to look again at the frames?
That is where it beat the rest: on Elorian's API it was off by 0.08 s on average and named frames between its samples. If an answer within about half a second is enough, Gemini was off by 0.4 s at USD 0.014 per clip.
Is sound part of the check?
It gets a transcript of the words, not the sound. It found 0 of 5 sound defects; free ffmpeg found 2.
Can you pay about 50 cents and 7 minutes per clip?
Our clips ran 20 to 36 seconds: about USD 0.018 per second of video. Cheaper per token than Gemini, about 36 times the price per clip.
How many false alarms can you take?
On our 3 clean clips it raised 3. Gemini and ffmpeg raised 2 each.
Checked
Sources
- Andrew Dai, launch post on X, 16 Oct 2026 checked on 29 Oct 2026
- Elorian, Introducing Glimpse, 16 Oct 2026 checked on 31 Oct 2026
- Elorian API docs checked on 31 Oct 2026
- OpenRouter docs, video input checked on 31 Oct 2026
- OpenRouter model list and prices checked on 31 Oct 2026
Yes, that's my AI avatar in the video. Same opinions, better lighting.