2026-08-03
AI transcription upload limits in 2026 (and why extracting audio beats uploading the whole video)
You drop a Zoom recording, a lecture capture, or a phone video into an AI transcription or chat tool, and it comes back with a file-too-large error, or a duration you didn't expect to get truncated at. The video usually isn't actually too long — it's carrying a video track's worth of bytes for content the tool only needed the audio from. Here are the specific, current caps three widely used tools enforce, and the one step — extracting just the audio track first — that clears most of them without losing anything the transcript needed in the first place.
The limits that actually apply right now
As of mid-2026:
- OpenAI's transcription API (Whisper, gpt-4o-transcribe, gpt-4o-mini-transcribe): 25MB, full stop. OpenAI's own docs are blunt about it: "Files can be up to 25 MB." That cap applies to the whole file, and the accepted formats list — mp3, mp4, mpeg, mpga, m4a, wav, webm — includes mp4 and webm as valid direct uploads to that same endpoint. Nothing stops you from handing it a video file; the 25MB budget just fills up far faster with one, because every second of video carries a video track's worth of bytes the transcription model never looks at.
- Otter.ai: a generous 5GB cap — but its own troubleshooting advice is to strip the video anyway. Otter's own help center puts the import ceiling at 5GB, which sounds like it shouldn't be a problem — but for import failures, its own fix is exactly this post's premise: "if you are importing a video file, export it as audio-only to reduce the file size." That's Otter telling its own users to do the video-to-audio step manually, not a workaround this post invented.
- Google Gemini apps, free tier: 5 minutes of video vs. 10 minutes of audio — literally double. Google's own Gemini Apps help center states it plainly: "Total video length can be up to 5 minutes," while "Total audio length can be up to 10 minutes." Upgrading to Google AI Pro or Ultra raises both ceilings — to 1 hour of video and 3 hours of audio — but the 2x gap between video and audio persists at every tier, not just the free one.
Why the same recording is so much smaller once it's audio-only
The gap isn't a quirk of any one tool — it's what each format actually has to encode. A video
track has to carry pixel data for every frame; an audio track only has to carry a waveform. NearVid's
audio extraction uses ffmpeg's libmp3lame encoder at -q:a 2, and
FFmpeg's own wiki documents exactly what that
setting produces: a VBR MP3 averaging 170-210 kilobits per second. Do the arithmetic on that
sourced range and a 25MB budget holds roughly 16 to 20 minutes of that audio — the exact number
depends on how compressible the specific recording is, but it's an estimate built on a documented
encoder setting, not a made-up multiplier. A video file at any typical recording bitrate — phone
video and screen recordings routinely run into multiple megabits per second, well above 170-210
kilobits — burns through the same 25MB budget far faster, which is the entire reason OpenAI's cap,
Gemini's duration limit, and Otter's own troubleshooting advice all point at the same fix.
What NearVid actually does here
To be specific about scope: NearVid's audio extraction pulls the entire audio track out of a
video file and saves it as an MP3 at that same fixed -q:a 2 setting — there's no
bitrate or quality picker for this particular output, unlike the High/Medium/Low presets WebM and
MP4 conversion offer. It also doesn't take a start/end range in that same step; if only part of a
long recording actually needs transcribing, the way to get there is NearVid's separate, lossless
trim first — cut to the relevant minutes, then run audio extraction on the trimmed file, two fast
steps instead of one. NearVid's own input ceiling (500MB per file on desktop, 50MB on mobile or
memory-constrained devices) is well above any of the three caps above, so a source recording large
enough to trip an AI tool's own limit still fits into NearVid to begin with. And it goes without
saying, but is worth being explicit about: NearVid doesn't transcribe anything itself — it produces
a smaller MP3 that then goes to whichever transcription or AI tool you were already using, entirely
on-device, with the original video file never leaving your browser tab to get there.
The honest summary
The numbers as of mid-2026: OpenAI's transcription API caps every upload at 25MB regardless of whether it's audio or video; Otter.ai allows 5GB but its own support docs recommend audio-only for import problems anyway; Gemini Apps' free tier allows exactly double the upload duration for audio versus video, a gap that persists even after upgrading to Pro or Ultra. None of those numbers move because of anything NearVid does — they're limits set by the tools you're uploading to. What changes is how quickly your file hits them: extracting the audio track first, at the same documented ~170-210kbit/s MP3 setting NearVid always uses, routinely buys back most of that budget before you've uploaded a single byte anywhere.