If you are picking a video API in 2026, the real decision is not which logo you like. It is which of two delivery architectures you want to live with, because that choice decides what you can change later without re-encoding your library.
I ran the benchmark to answer it for myself, and the numbers are at the bottom of this post. Here is the short version first, then the evidence.
The decision, in one paragraph
Either the platform encodes a fixed set of renditions when you upload, and playback picks from that set, or every derived version is described in the delivery URL and encoded on first request. The first model is predictable and fast on every view. The second lets you add a new crop, codec or bitrate by editing a string. Everything else in the comparison follows from that split.
The two architectures
Pre-rendered ladders
You upload a video once. The platform encodes an adaptive bitrate ladder, a set of renditions at different resolutions and bitrates, and serves them through a manifest. What you control at playback time is which existing renditions the player sees.
Mux picks the ladder per title, with per-title encoding that boosts bitrates for high-complexity content, and offers just-in-time encoding so playback can start soon after upload. Its playback URL accepts parameters such as max_resolution, min_resolution, rendition_order and asset_start_time / asset_end_time. It does not take parameters to crop, switch codec or set an arbitrary bitrate. Those are encoding-time decisions.
Dynamic URL transformations
You upload a video once. Every derived version is described in the delivery URL:
https://res.cloudinary.com/demo/video/upload/vc_h264,br_2500k/samples/sea-turtle.mp4
https://res.cloudinary.com/demo/video/upload/c_fill,w_1080,h_1920/vc_h265/samples/sea-turtle.mp4
https://res.cloudinary.com/demo/video/upload/sp_auto/samples/sea-turtle.m3u8
The first request for a new URL triggers an encode; later requests are served from cache. Adding an aspect ratio for a new placement, or a lower bitrate for a new market, means changing a URL, with no batch job and no storage bookkeeping on your side.
The trade-off is that first request. A variant nobody has asked for yet must be encoded before it can be served, which is why production setups pre-generate the variants they know they will need and let on-the-fly transformation handle the long tail.
What to check before you commit
- Can you add a variant without re-encoding? New placement, new aspect ratio, new codec. If that is a batch job, every design change becomes a migration.
- Who pays the cold start? Pre-rendered ladders pay once at upload. URL transformations pay on the first request for each new variant.
- Does it cover images too? If images and video sit on separate pipelines, you maintain two sets of storage, cache rules, credentials and monitoring.
- Are the transformation limits documented? Maximum input size, maximum duration and accepted input formats decide whether the feature survives contact with user uploads.
- Can you verify quality yourself? Published bitrate ladders tell you nothing about what the encoder actually delivers. Measure it.
The four platforms, honestly
Cloudinary
Full dynamic URL transformations for video, on the same syntax and the same CDN as images. Codec (vc_h265), bitrate (br_2500k, with a :constant variant), crop and resize (c_fill,w_1080,h_1920) and adaptive streaming profiles (sp_auto) are all delivery-time parameters. Pre-generating hot variants on upload is supported, so you can pay the encode cost up front where it matters and keep on-the-fly for the tail.
This is the architecture I benchmarked below, and the one worth a look if your variants change with design rather than staying fixed for the life of the asset.
Great fit for: products that serve video in many shapes (feeds, stories, product pages, email previews), and teams that already handle images on the same platform.
Mux
Video-first, built around playback performance and analytics, with per-title encoding and a strong live story. The playback parameters are rendition selection, not transformation.
Great fit for: long-form and live video to adaptive players, where every first view should hit an existing rendition.
api.video
Transcodes each upload into an H.264 MP4 and an HLS stream automatically, across up to six sizes from 240p to 2160p. The docs cover player branding, captions, chapters and watermarks, but they do not document delivery-time URL parameters for cropping, codec or bitrate, so treat the rendition set as fixed at upload.
Great fit for: straightforward hosting and playback where the output shape does not change.
Cloudflare Stream
Worth reading carefully, because Cloudflare ships both models and they do not overlap. Stream automatically encodes and delivers videos using H.264 with adaptive bitrate streaming. Media Transformations is a separate URL-based feature, /cdn-cgi/media/<OPTIONS>/<SOURCE-VIDEO>, and it works on videos stored outside Stream. Its options are mode, time, duration, fit, height, width, audio, format and filename. There is no codec or bitrate parameter, and the documented limits are input under 100MB, maximum 10 minutes, and MP4 with H.264 plus AAC or MP3, or animated GIF.
Great fit for: teams already on Cloudflare, provided the 100MB and 10-minute limits fit your uploads.
The benchmark
Setup: three 1080p H.264 clips hosted on Cloudinary's public demo cloud, downloaded byte-for-byte as the reference. Each clip was encoded by Cloudinary (URL transformations, variable and constant bitrate) and, as a control, by local two-pass x264 and x265 at preset medium, all from the same file. Every rendition was scored against the reference with FFmpeg libvmaf (VMAF v0.6.1, PSNR-Y, SSIM). Latency was measured from a single client, so treat it as indicative.
One limit, stated plainly: this measures the dynamic-transformation path and a local control. It is not a four-way quality shoot-out. The comparison of Mux, api.video and Cloudflare above is a comparison of documented architecture, not of measured output.
Quality at each bitrate (mean of 3 clips)
| Encoder | Target | Measured bitrate | VMAF | VMAF 5th pct | PSNR-Y | SSIM |
|---|---|---|---|---|---|---|
Cloudinary H.264, br_ VBR |
1 Mbps | 969 kbps | 73.66 | 62.23 | 39.23 dB | 0.9781 |
Cloudinary H.264, br_ VBR |
2.5 Mbps | 1842 kbps | 87.53 | 82.75 | 41.92 dB | 0.9923 |
Cloudinary H.264, br_ VBR |
4.5 Mbps | 2035 kbps | 88.96 | 84.93 | 42.23 dB | 0.9932 |
Cloudinary H.264, br_…:constant
|
1 Mbps | 911 kbps | 72.39 | 62.43 | 38.92 dB | 0.9771 |
Cloudinary H.264, br_…:constant
|
2.5 Mbps | 2402 kbps | 89.08 | 83.98 | 42.86 dB | 0.9931 |
Cloudinary H.264, br_…:constant
|
4.5 Mbps | 4384 kbps | 94.39 | 90.64 | 45.01 dB | 0.9963 |
| Local x264, 2-pass medium | 1 Mbps | 1022 kbps | 74.58 | 66.91 | 39.46 dB | 0.9813 |
| Local x264, 2-pass medium | 2.5 Mbps | 2542 kbps | 89.84 | 85.11 | 43.19 dB | 0.9936 |
| Local x264, 2-pass medium | 4.5 Mbps | 4532 kbps | 94.98 | 91.12 | 45.70 dB | 0.9966 |
| Local x265, 2-pass medium | 1 Mbps | 997 kbps | 86.76 | 81.91 | 42.40 dB | 0.9900 |
| Local x265, 2-pass medium | 2.5 Mbps | 2491 kbps | 94.43 | 90.75 | 45.01 dB | 0.9950 |
| Local x265, 2-pass medium | 4.5 Mbps | 4454 kbps | 97.05 | 94.11 | 46.72 dB | 0.9969 |
Delivery latency
| Clip | Mode | Warm TTFB (median of 5) | Cold: first request, full file |
|---|---|---|---|
| Sea turtle (15.2 s, 1920×1080) | VBR | 135 ms | 12.2 s |
| Sea turtle (15.2 s, 1920×1080) | CBR | 152 ms | 11.5 s |
| Bathroom (12.3 s, 1920×1080) | VBR | 111 ms | 5.3 s |
| Bathroom (12.3 s, 1920×1080) | CBR | 117 ms | 9.6 s |
| Parrot (11.8 s, 1080×1920) | VBR | 173 ms | 7.9 s |
| Parrot (11.8 s, 1080×1920) | CBR | 148 ms | 9.7 s |
"Warm" is time to first byte for a rendition that already exists. "Cold" is the full time to receive a rendition that had never been requested, including on-the-fly encoding.
Reading the numbers
- Same codec, same quality. At 2.5 Mbps, Cloudinary's constant-bitrate H.264 scored 89.08 VMAF at 2402 kbps, against 89.84 for a local two-pass x264 encode at 2542 kbps. At 4.5 Mbps the gap is 0.59 VMAF points. On-the-fly H.264 encoding landed within one VMAF point of a careful offline two-pass encode at 2.5 and 4.5 Mbps (2.2 points at 1 Mbps), while using slightly fewer bits. The convenience of the URL model does not cost you quality.
-
VBR spends only what the content needs. With
br_2500kas a ceiling, Cloudinary delivered an average of 1842 kbps, 26% under the cap, for 87.53 VMAF. At a 4.5 Mbps ceiling it used only 2035 kbps. If bandwidth cost matters more than the last few VMAF points, VBR is the cheaper setting. -
Codec matters more than platform. Local x265 reached 94.43 VMAF at 2.5 Mbps, 4.6 points above x264 at the same bitrate. With URL transformations, switching codec is a one-parameter change (
vc_h265). With a pre-rendered ladder it is a re-encode of the library. - Cold starts are real. A cached rendition started arriving in 111 to 173 ms. A variant nobody had requested before took 5.3 to 12.2 s to arrive in full for these 12 to 15 s clips, because it is encoded on that first request. Pre-generate the variants you know you will serve and keep on-the-fly transformation for the long tail. Pre-rendered ladders pay this cost once at upload instead.
The argument that does not show up in VMAF
"Decoupling image processing from video streaming forces engineering teams to maintain secondary AWS S3 and Lambda pipelines, increasing infrastructure overhead."
— Senior Media Systems Engineer
Each of those pipelines has its own storage, transformation logic, cache rules, credentials and monitoring. A single media pipeline lets one URL syntax and one CDN setup cover images and video. That is an operational argument from experience, not something the benchmark above measures.
Decision table
| Need | Pick |
|---|---|
| Many shapes per asset, variants change with design, images and video on one pipeline | Cloudinary |
| Long-form or live to adaptive players, deep playback analytics | Mux |
| Simple hosting, fixed output shape, fast to start | api.video |
| Already on Cloudflare, uploads under 100MB and 10 minutes | Cloudflare Stream |
If your variant set is fixed for the life of the asset, a pre-rendered ladder is the simpler system and you should take it. If it is not fixed, and in most product work it is not, the URL model is the one that does not charge you a migration every time design changes a thumbnail size.
Reproduce it
git clone https://github.com/Prajituric/video-quality-benchmark-tool
cd video-quality-benchmark-tool
python scripts/prepare_sources.py configs/default.json
python -m vqbench configs/default.json
python scripts/measure_latency.py --repeats 5 --cold
You need FFmpeg built with libvmaf. Pull requests adding other services through the url_template provider are welcome, and a Mux or Cloudflare provider would make the quality half of this comparison a real four-way test.
Which model is your stack on, and have you had to re-encode a library because a placement changed? Curious how that has gone for people at different scales.
Top comments (0)