The real AI transcription limits in 2026: every major tool's hard ceiling

Ask ChatGPT to transcribe a two-hour interview and it will try, then fail. Ask Claude and it won't even accept the file. This is not a bug you can work around with a better prompt. Every general-purpose assistant treats audio as a bolt-on feature with a hard ceiling, and once you know each ceiling's exact number, the pattern is impossible to unsee.

The reason isn't laziness. It's architecture, and we'll get to the specific 30-second constraint underneath all of it. First, the numbers - the ones your tools don't advertise on the pricing page. What follows is the real 2026 limit for each major tool, pulled from vendor documentation, so you can find your own answer to "can this thing handle my recording?" before you're an hour in and it quits.

One caveat, stated up front: these are moving targets. ChatGPT Record, Gemini audio, and NotebookLM's tiers all shipped within the last year, and every figure here should be read as a snapshot. Last verified: 2026-07-12. If you're reading this much later, treat the numbers as a starting point and re-check the linked source.

OpenAI's transcription API caps every upload at 25 MB, across all supported formats (mp3, mp4, mpeg, mpga, m4a, wav, webm). That single number is the wall behind "ChatGPT can't take my file." Anything larger has to be compressed or split into chunks before it goes anywhere. For a typical 128 kbps MP3 that works out to roughly 25 minutes of audio - an approximation, since the documented limit is the 25 MB size, not a duration.

There's a second, quieter limit that trips people up: format and features. Among OpenAI's transcription models, only whisper-1 returns word-level timestamps and SRT, VTT, or verbose JSON output. The newer gpt-4o-transcribe and gpt-4o-mini-transcribe return only plain text or JSON - no caption files. And whisper-1 cannot label speakers at all; diarization is supported only by the separate gpt-4o-transcribe-diarize model. So the one model that gives you subtitles is not the one that gives you speaker turns.

None of this makes ChatGPT useless for a short clip. It makes it the wrong tool for a long, multi-track recording. If you want the full walkthrough, we cover whether ChatGPT can transcribe audio and how to get past the 25 MB limit in the guides.

Claude's own "Upload files to Claude" help page lists exactly two categories of accepted upload: documents (PDF, DOCX, CSV, TXT, HTML, ODT, RTF, EPUB, JSON, XLSX) and images (JPEG, PNG, GIF, WebP). Audio and video are not on the list. You cannot drop an MP3 or an M4A into the Claude web or desktop app and get a transcript back.

This isn't an oversight to route around - it's how the product works. Anthropic's models don't take native audio input, so any workflow you've seen described as "Claude transcription" means the audio was transcribed somewhere else first and the resulting text was pasted in. Claude is excellent at cleaning up, summarizing, or restructuring a transcript you already have. It is simply not the thing that turns sound into words.

Google added audio uploads to the Gemini app on 8 September 2025, with a hard duration ceiling that most free users hit fast: up to 10 minutes of audio total per prompt. That budget is cumulative and shared across up to 10 attached files in a single prompt, so a couple of voice memos can exhaust one turn's allowance.

Paying lifts the ceiling considerably. Google AI Pro and AI Ultra subscribers can upload up to 3 hours of audio. But raising the limit doesn't change the output: the Gemini app returns plain analysis, not a captioned SRT/VTT file or a diarized transcript. A three-hour interview still comes back as prose you have to prompt for and manually stitch together, with no timestamps and no speaker labels. If subtitles are the goal, a dedicated subtitle generator is a different kind of tool.

One honesty note on this figure: the exact caps come from Google's September 2025 announcement as reported by multiple outlets rather than a standing help-center page that enumerates the minutes, so treat them as well-sourced but subject to change.

NotebookLM is the interesting case, because it does accept audio - MP3, WAV, M4A and more - and transcribes it on import. Free users get 50 sources per notebook, each source capped at 500,000 words or 200 MB. On paper that sounds generous.

The catch is what you get back. NotebookLM is a grounded-research tool, not a transcription export tool. It will happily answer questions about your recording, but the transcribed text has no timestamps, no speaker labels, and no way to download it as an SRT, VTT, or DOCX transcript. Your two-hour interview becomes a searchable blob you can query and never a document you can hand off. (That absence is observed product behavior rather than a stated limit, so verify it against your own notebook if it's load-bearing for you.)

For anyone whose next step is a captions track, a quote-checked article, or a file for a coding tool, that's a dead end - the same complaint we heard again and again when we read 100 Reddit threads about transcription.

Purpose-built meeting tools do better than the general assistants, but their free and entry tiers still cap out. Otter.ai's Basic (free) plan allows 300 transcription minutes per month, 30 minutes per single conversation, and just 3 lifetime audio or video file imports. A single long recording blows the per-conversation limit; a handful of uploads exhausts the imports for good. (Otter has changed the import figure historically, which is exactly why the last-verified date matters here.)

ChatGPT Record is the other meeting entrant, and it's narrower than it looks. It runs only in the macOS desktop app, is gated to Plus, Pro, Business, Enterprise and Edu workspaces, and produces an editable summary with time-stamped citations and action items - not a raw diarized transcript you own. Third-party reviews report a per-session cap of roughly 2 hours (120 minutes), but that number is not confirmed in OpenAI's own documentation, so read it as reported rather than official.

If the appeal of these tools is "I only need this occasionally," the honest fit is a pay-as-you-go tool, not a recurring plan you use in bursts - the case we make in transcription without a subscription. And if you mainly want the gist, a meeting summarizer is a legitimately good answer.

Here's the architecture that explains every limit above. Whisper - the OpenAI speech model that sits under most of these tools - was trained on 30-second audio chunks, and the paper states plainly that the models "cannot consume longer audio inputs at once." Long audio has to be sliced into 30-second segments and re-stitched using predicted timestamps. That single constraint is the underlying reason every consumer tool imposes a chunk or size limit somewhere.

It's worth saying that Whisper is genuinely good at what it's designed for. Published benchmarks show it scoring 2.7% word error rate on LibriSpeech test-clean - near-human accuracy on clean, read speech - after training on 680,000 hours of audio, with a 55.2% average relative error reduction versus supervised models when tested zero-shot on out-of-distribution data. If you want the plain-English version of that metric, we explain word error rate and how accurate AI transcription really is.

The 2.7% number is the best case. The same Whisper paper shows the accuracy falling off a cliff the moment you leave clean read speech: 9.0% WER on Common Voice, 13.8% on Switchboard, 17.6% on CallHome telephone speech, and 36.4% on AMI SDM1 - far-field, multi-speaker meeting audio. A real meeting recorded on a laptop mic across a table is much closer to that last number than the first. That gap is where a free tool goes from "impressive" to "why is every third word wrong," and it's mostly fixable upstream, which is what improving transcription accuracy is about.

Then there's the volume math. English conversation runs roughly 120 to 150 words per minute, so a two-hour interview is somewhere near 18,000 words. Set that against the ceilings above: it sails past ChatGPT's 25 MB (about 25-minute) wall, laps Gemini free's 10-minute-per-prompt cap twelve times over, and exceeds Otter Basic's 30-minute-per-conversation limit fourfold. The general tools aren't a little short for the long job. They're an order of magnitude short.

For a one-off short clip - a two-minute voice memo, a quick quote to check - the honest answer is that a free tool is the right tool. ChatGPT will handle it, your phone's built-in dictation will handle it, and there's no reason to pay for anything. Reach for the free option and don't overthink it.

The tools break down on a specific job: long, multi-speaker, export-heavy audio. A full interview, a deposition, a season of podcast episodes, a research corpus that has to come out as real files. That's the shape Pepys was built for - no per-file length cap, so a two-hour recording doesn't die at the 30-minute mark; real exports (TXT, SRT, VTT, DOCX, PDF, JSON) instead of a blob you can only query; and no training on your audio. It appears here at the end, not the top, because for anything a free tool can already do, you should use the free tool. Pepys is the answer for the recording that made you read this far.