A data report from RedPanda Compress. All statistics are aggregates over tens of thousands of real-world video files analyzed in August 2026, with a September re-run added below — percent-only, no user-level data. Reuse welcome with attribution (see the end of the post).
September update: we re-ran everything — what moved, what didn’t
In September we re-ran every query on a fresh sample, with identical definitions so the two months compare like for like. The headline is how little changed: the shape of everyday video looks the same in two independent samples. The findings below still describe August; this section is the delta.
| Metric | August | September |
|---|---|---|
| H.264 share of files | 82.1% | 80.5% |
| HEVC share of files | 13.2% | 14.8% |
| AV1 share of files | 0.22% | 0.26% |
| MPEG-4 Part 2 files per AV1 file | 6.3 | 5.9 |
| .mp4 files that aren’t H.264 | 14.8% | 16.5% |
| Files longer than 1 hour | 7.8% | 7.3% |
| Desktop video that is portrait | 23.2% | 23.5% |
| Files under 2 Mbps (already compressed once) | 24.9% | 25.8% |
| Files above 20 Mbps | 15.1% | 14.8% |
| Compression jobs on ≤3G-class connections | 9.8% | 10.0% |
| Desktop files last touched by FFmpeg | 41.7% | 42.3% |
| Desktop files still carrying a camera fingerprint | 11.8% | 11.0% |
What moved: HEVC keeps creeping up, and the “MP4 that isn’t H.264” is now one in six. H.264 lost 1.7 points and HEVC gained 1.6. The same shift shows up in desktop-only files (H.264 81.6% → 79.8%), so it isn’t just a change in the mix of devices. Inside .mp4 files, the non-H.264 share went from 14.8% to 16.5% (HEVC alone: 13.3%), so Finding 7’s “one in seven” is now closer to “one in six”. Across the September sample HEVC holds steady around 15% rather than climbing, so we read this as a drift toward a plateau, not a cliff. AV1 ticked up to 0.26% — still invisible, still outnumbered about six to one by the DivX-era codec.
What stayed put. Duration, orientation, bitrate, connection quality and FFmpeg’s share of the last-writer fingerprints all moved by roughly a point or less. Two independent samples landing on the same shape is the best evidence we have that these are properties of everyday video, not artifacts of one month’s traffic. The modal bitrate is still 5–10 Mbps (22.0%, was 23.2%), and about one file in four is still a re-compression of already-compressed video.
What faded: the sticker cluster. Square video is now 4.3% of desktop files (was 6.3%), and 44% of all square clips are below 360p (was 61%). The August “chat-sticker culture” cluster was real but not permanent — a reminder to read small, tight clusters in usage data as snapshots rather than trends.
Method notes for the update: orientation and resolution could not be determined for a larger share of desktop files in September (about 47%, up from about 38%), so every orientation figure remains known-only. Encoder-fingerprint rows use desktop files only, where the fingerprint could be read for about 98% of files in both months (the all-device August figure in Finding 6 was 41.0%). The mix of devices in our sample shifts from month to month, so where a comparison could be distorted by it we checked the desktop-only cut.
Methodology & privacy, up front
RedPanda Compress is a browser-based video compressor: files are processed entirely on the user’s device and never uploaded. The telemetry behind this report is deliberately coarse — pre-bucketed ranges (resolution class, duration range, bitrate range), codec names, and connection-quality classes. No filenames, no file contents, no metadata atoms (camera model, GPS), and no per-user profiles; the categories match our public privacy page.
Known biases: this is not a census of “all video” — it is a census of video people need to shrink, so heavy formats are overrepresented. Orientation and resolution could not be determined for roughly a third of files; those are excluded from the orientation and resolution figures. Connection classes come from the browser’s Network Information API, which estimates effective quality — “3G” means “performs like 3G,” not a cell contract. Encoder provenance was measured on a large subsample.
Finding 1 — H.264 is still 82% of everything. AV1 is statistically invisible.
| Codec | Share of files |
|---|---|
| H.264/AVC | 82.1% |
| HEVC/H.265 | 13.2% |
| MJPEG | 1.6% |
| MPEG-4 Part 2 (the DivX/Xvid era) | 1.4% |
| ProRes | 0.23% |
| AV1 | 0.22% |
| VP9 | 0.18% |
| VP8 | 0.08% |
Twenty-three years after standardization, H.264 remains the water everyone swims in. The striking pair: files encoded with ~2003-era MPEG-4 Part 2 still outnumber AV1 files six to one. AV1 has won the streaming-platform war — YouTube and Netflix serve it billions of times a day — but in the world of files people actually hold (camera output, exports, downloads, old archives) it has essentially no presence. Codec adoption in personal files is generational, not technological: files outlive the codecs that made them, and the encoder defaults of cameras and apps — not the preferences of standards bodies — decide what the world’s disks look like. The one codec visibly gaining is HEVC, and it is gaining the same way H.264 did: by being a phone camera default.
Finding 2 — “Video” no longer means “clip”
| Duration | Share |
|---|---|
| < 15 s | 11.8% |
| 15–60 s | 22.5% |
| 1–5 min | 30.1% |
| 5–20 min | 17.1% |
| 20–60 min | 10.5% |
| ≥ 1 hour | 7.8% |
Nearly one file in five runs longer than 20 minutes, and roughly one in thirteen exceeds an hour — lectures, meetings, screen recordings, gameplay sessions. The mental model of compression as “shrink this phone clip” misses a fifth of the real workload: video is now also a document format, the recording of something that took an hour because the thing itself took an hour.
Finding 3 — Nearly a quarter of desktop video is vertical
Orientation of files processed on desktop computers (where orientation could be determined):
| Orientation | Share of desktop files |
|---|---|
| Landscape | 70.5% |
| Portrait | 23.2% |
| Square | 6.3% |
Vertical video was born on phones, but it now flows routinely through desktop workflows — footage synced, transferred, or downloaded to a PC for editing and sharing. Desktop software that treats portrait video as an edge case is failing nearly a quarter of real files.
Inside the square segment hides this report’s favorite curiosity: 61% of all square videos are below 360p, and 71% run under 15 seconds — a distinct, many-user cluster consistent with animated chat-sticker and emote culture. Even a three-second looping sticker is worth compressing when a chat app enforces a size cap.
Finding 4 — The 5–10 Mbps world, and the quarter that’s already been compressed
Container bitrate (size × 8 ÷ duration) of source files:
| Bitrate | Share |
|---|---|
| < 1 Mbps | 10.7% |
| 1–2 Mbps | 14.2% |
| 2–5 Mbps | 16.1% |
| 5–10 Mbps | 23.2% (modal) |
| 10–20 Mbps | 20.4% |
| 20–50 Mbps | 9.3% |
| ≥ 50 Mbps | 5.8% |
The modal file arrives at 5–10 Mbps — the default output of phone cameras and screen recorders. But the tails tell the story. One file in seven exceeds 20 Mbps — modern phones shooting high-bitrate 4K their owners immediately need to shrink. And a full quarter of files arrive below 2 Mbps: video that has already been compressed once, being compressed again to squeeze under some app’s attachment limit. Both tails are artifacts of the same mismatch — recording defaults and sharing limits are set by different companies, and users are stuck reconciling them.
Finding 5 — The users the cloud forgets
Roughly one compression job in ten starts on a connection the browser classifies as 3G-class or slower (per the Network Information API’s effective-quality estimate); on desktop alone the share is about one in twelve. The label surprises until you remember what the API measures: effective throughput, which sweeps congested Wi-Fi, VPNs, tethering, and ISP throttling into the same bucket as genuine 3G.
For this population, uploading a gigabyte of video to a cloud service is somewhere between painful and impossible. Client-side processing isn’t a privacy preference for them; it is the only version of the product that works at all.
Finding 6 — 41% of everyday video has been touched by FFmpeg
Video files carry faint fingerprints of the last software that wrote them. Classifying those fingerprints on a large subsample:
| Last writer | Share |
|---|---|
| FFmpeg-family (Lavf muxer) | 41.0% |
| Other/indeterminate tools | 28.8% |
| Video editors | 12.4% |
| Android phone cameras | 7.7% |
| Unclassifiable | 4.2% |
| Apple cameras | 2.8% |
| Platform downloads (YouTube-style) | 2.2% |
| Action cameras | 1.0% |
Two things stand out. First, FFmpeg — one open-source project — was the last tool to touch two files in five. It is the invisible plumbing inside converters, downloaders, editors, transcoding pipelines and apps that never mention it; no other single piece of software comes close.
Second, read the camera rows together: only about one file in nine still carries a camera’s own fingerprint. Everything else has already been through at least one piece of software — trimmed, converted, downloaded, re-muxed — before reaching us. The “original” straight-off-the-camera video is, by the time anyone needs to share it, a minority artifact.
Honest limits: the fingerprint only names the last writer, not the chain; “Lavf” is a giant catch-all for anything built on FFmpeg; and a missing fingerprint does not prove a file came straight from a camera.
Finding 7 — .mp4 is a monoculture, and one MP4 in seven is lying to you
By file extension, everyday video is astonishingly uniform: nine files in ten are named .mp4. MOV takes most of the rest (7%); MKV, AVI and WebM are all below 1% each. The container war is over.
But the label has quietly stopped meaning what people think it means:
What’s actually inside a .mp4 | Share |
|---|---|
| H.264 (plays everywhere) | 85.2% |
| HEVC | 11.6% |
| MPEG-4 Part 2 | 1.3% |
| MJPEG | 1.2% |
| AV1 + others | 0.7% |
14.8% of .mp4 files — one in seven — don’t contain H.264, and most of those carry HEVC, which still fails to play in many browsers and on many non-Apple devices. “It’s an MP4, it’ll play anywhere” was true for fifteen years; phone cameras defaulting to HEVC-in-mp4 have silently broken it. The extension names the box, not the contents — and the box is no longer a guarantee.
(A small aside for trivia lovers: 3% of files have no audio track at all — mostly screen recordings and camera timelapses.)
Summary for the impatient
- H.264 is 82% of personal video; AV1 is 0.22% — still outnumbered 6:1 by DivX-era MPEG-4.
- Nearly 1 in 5 files exceeds 20 minutes — compression is now a meetings-and-lectures workload.
- 23% of desktop video is vertical — and 61% of square video is sub-360p chat stickers.
- The modal file is 5–10 Mbps, but a quarter of files are re-compressions of already-compressed video.
- ~10% of compression jobs start on ≤3G-class connections — the population client-side tools exist for.
- FFmpeg was the last tool to touch 41% of files; only ~1 in 9 still carries a camera fingerprint.
- 90% of files are named .mp4 — but one MP4 in seven doesn’t contain H.264 and may not play everywhere.
Reuse this data
Data and charts may be reused with attribution and a link to redpandacompress.com. Questions about methodology: [email protected]. All statistics are aggregates over coarse buckets; no user-level data exists to share. If you’re curious how a browser can analyze and compress video without uploading it, we’ve written up how the in-browser pipeline works and how to verify no upload happens.
Fei is a skilled software engineer. He previously worked at Google and now at a startup. His expertise includes web media processing, cloud architecture, complex algorithms, and AI training and deployment. Beyond work, Fei enjoys diving into new knowledge and is a big fan of strategy games.
