OmniHuman 1.5 vs OmniHuman 2.0: A Third Cheaper, Marked Superseded, and What We Cannot Yet Say About Quality (October 2026)
fal lists OmniHuman 2.0 at $0.105 per second against $0.16 for 1.5, and marks 1.5 as superseded. What that means for cost, which of your workflows can move, and why you should still run a paired test before you switch.
If you built a talking-photo workflow on OmniHuman 1.5, fal has changed the ground under you. In mid-October 2026 it added OmniHuman 2.0 at $0.105 per second, marked the 1.5 endpoint as superseded by it, and left 1.5 callable at $0.16. Superseded on fal means the platform points new work at the newer model; it does not by itself mean the old one is switched off, and it does not tell you the new one is better for your clip. Here is what is documented and what has to be tested.
The side-by-side
| OmniHuman 1.5 | OmniHuman 2.0 | |
|---|---|---|
| Price per second of audio | $0.16 | $0.105 |
| Cost of 15 seconds | $2.40 | $1.58 |
| Cost of 1 minute | $9.60 | $6.30 |
| Status on fal | superseded | current |
| Audio length | not restated here | 4 to 300 seconds |
| Image aspect ratio | not restated here | 0.4 to 2.5 |
| Resolution setting | not restated here | none |
The saving is 34 percent on every clip, because both models bill purely by the length of the audio. For a creator making twenty 15-second talking clips a month, that is $16 saved; for a service making thousands, it is the reason to migrate.
What is the same
Both take a photo and an audio file and make the person in the photo perform the audio, with expressions and upper-body movement following the waveform rather than a text prompt. Both are ByteDance models. Both accepted real people's photos when we and others used 1.5 on fal. And both inherit the same practical rules: a frontal, evenly lit photo; short clips for the best identity; a restrained prompt if you use one. If you are new to the family, the OmniHuman 2.0 guide covers inputs and limits.
What changed that we can document
- Price: down by about a third.
- Label: 2.0 is the recommended endpoint on fal, 1.5 is not.
- Documentation: ByteDance updated the model's API documentation on 13 to 15 October, which is how the release date was dated, since fal's listing does not carry one.
What we cannot tell you
Whether 2.0 looks better, worse or the same. No one has published a controlled comparison we trust, and we have not yet run 2.0 on our own material. The version number suggests an improvement and the lower price suggests a more efficient model, but both are inferences. For a task where the face is the point, an inference is not enough.
Whether the failures are the same. OmniHuman on fal was unavailable on 16 July and again on 4 September, which broke a test of ours on availability rather than on policy. A newer endpoint could be more stable or less; there is no record yet.
How to migrate without regret
- Pick two clips you already made on 1.5: one easy (calm, frontal) and one hard (a lively delivery, an unusual angle).
- Run the same photo and audio through 2.0 and watch both versions at full screen on a phone, with sound. Look at the mouth on plosives, at the eyes in the pauses, and at the last second of the clip.
- If 2.0 matches or beats 1.5, switch. At about $1.05 to $1.58 per test clip, the experiment costs less than one retry.
- Keep 1.5 as a fallback for as long as fal serves it; a superseded model is a useful second source when the new one has an outage.
If you use a tool rather than the API, ask it which version it runs and whether the price changed. Our own Studio's model catalogue still lists OmniHuman 1.5 while we test 2.0, so the credit price you see there reflects 1.5 until that changes.
Which one if you are choosing today
For a new project, start with 2.0: it is cheaper, current and the one fal points to. Use 1.5 only where an integration is pinned to it. If neither suits, Kling AI Avatar v2 Pro at $0.115 and MiniMax H3 Max Lip Sync at a $0.05 base for long audio are the nearest alternatives, compared in the talking-photo roundup.
Common questions
Will 1.5 stop working?
Superseded on fal means the platform points new work at 2.0. It does not announce a shutdown date, but a pinned integration should have a plan.
Is 2.0 better?
It is cheaper and current. Whether the output is better for your clips is something only a paired test can say.
How much will I save?
About 34 percent per clip, because both versions bill only by audio length: $0.105 against $0.16 per second.
Can I run both?
Yes. Run 2.0 as the default and keep 1.5 as a fallback while fal still serves it.
How to do this without code: You do not manage model versions in the product: the talking-photo path uses what is in the catalogue and quotes the price. In AutoHustle this is the Animate photo tool with a voice recording: audio in, a talking photo out (about 3 credits per second with the audio-driven model, quoted before you press Generate). See How to make a talking photo and Animate a photo.
Related reading: HappyHorse vs OmniHuman: Script-Driven or Audio-Driven Talking Video · AI Video Models on fal in October 2026: What Is New, What It Costs, and What Is Gone
For developers: the fal endpoints
| Model | fal endpoint |
|---|---|
| OmniHuman 2.0 | bytedance/omnihuman/v2 |
| OmniHuman 1.5 (superseded) | fal-ai/bytedance/omnihuman/v1.5 |
Prices read from fal on 18 October 2026: 2.0 at $0.105 per second, 1.5 at $0.16 per second. Cost is a function of audio length only.