FlashTalk on fal: A $0.02-per-Second Talking Head That Cannot Make a Vertical Video (October 2026)
SoulX-FlashTalk is advertised as the cheapest talking-head model on fal at $0.02 per second, but its output is fixed at 768 by 448 pixels, a horizontal frame. Why that rules it out for Reels, and where it is still a fair buy.
At first glance FlashTalk is the bargain of the talking-photo category. SoulX-FlashTalk is a 14-billion-parameter model that fal names as its published pick for talking-head generation, and it lists at $0.02 per second. For comparison: OmniHuman 2.0 is $0.105, Kling AI Avatar v2 Pro $0.115 and Standard $0.0562. A one-minute talking head would cost $1.20. The catch is in the output format, and it is a decisive one for anyone making short-form vertical video.
The catch: 768 by 448
The schema fixes the output at 768 by 448 pixels at 25 frames per second. That is a horizontal frame (16:9 and a bit), with no option to change the size or the aspect ratio. We checked it twice, once in September and again on 18 October, because it is the kind of constraint that is easy to miss on a pricing page. There is no 9:16 mode, and cropping a 768 by 448 landscape frame to vertical leaves a strip of about 252 by 448 pixels, which is smaller than a phone's thumbnail grid and unusable as a Reel.
| FlashTalk | OmniHuman 2.0 | Kling Avatar v2 Standard | |
|---|---|---|---|
| Price per second | $0.02 | $0.105 | $0.0562 |
| 1 minute | $1.20 | $6.30 | $3.37 |
| Output shape | fixed 768×448 landscape | follows the photo's aspect | follows the photo's aspect |
| Vertical (9:16) possible | no | yes | yes |
So when is it a fair buy?
Landscape is not wrong for everything. FlashTalk can make sense when:
- the destination is horizontal: a YouTube video, a website hero, an embedded explainer, a slide presentation;
- the face is small in the frame: a presenter in a corner of a screen recording, where 768 by 448 is plenty;
- volume matters more than polish: internal training, drafts of a long script, a prototype to judge a voice before spending on a premium model;
- you are testing speech timing and want to see the lip sync cheaply before committing.
For vertical social video, it is a trap: the price invites a try, the output arrives horizontal, and the money is spent. This is the single most important thing to know about it, and why we wrote the piece.
What we know and do not know
The price, the fixed size and the frame rate come from fal's schema. The 14-billion-parameter size and the "published pick" description come from fal's own listing. We have not generated with FlashTalk, so we cannot say how its lip sync, its head movement or its handling of real faces compare with the models above. As with every model in this category, no one has certified it for identity.
What to use for a vertical talking head
| You have | Use |
|---|---|
| A recording and a small budget | Kling AI Avatar v2 Standard at $0.0562 |
| A recording and a long length | MiniMax H3 Max Lip Sync (a $0.05 base, up to 15 minutes) |
| A recording and delivery that matters | OmniHuman 2.0 |
| A script and no recording | HappyHorse 1.1 image-to-video |
Our broader roundup of the category is talking photo apps compared.
A general lesson about cheap models
The cheapest option in a category is often cheap because it solves a narrower problem. Before you compare a price, compare the output: its shape, its length limit, its resolution, and whether the thing you intend to make is possible at all. A model that costs a fifth as much and cannot produce your format costs infinitely more per usable clip. The same discipline applies to base-rate pricing with unpublished resolution multipliers, which we flag in the Kling 4K guide, and to endpoints whose price unit is not explained, like LTX 2.5.
A quick way to confirm the size yourself
If you want to see the constraint rather than take our word for it, run a three-second clip and inspect the file: it will report 768 by 448 pixels at 25 frames per second. That costs about six cents and saves you from building a workflow around an output shape you cannot post. The same habit works for any new endpoint: before you plan around a price, generate the smallest possible clip and read the metadata of the file that comes back. Resolution, frame rate, duration and aspect ratio tell you more about fit than any pricing page.
Common questions
Can I crop the output to vertical?
A 768 by 448 frame cropped to 9:16 leaves about 252 by 448 pixels, too small for a Reel.
Is it good for lip sync?
We have not generated with it, so we cannot rate its lip sync. The price and the fixed size come from the schema.
What is it for, then?
Horizontal destinations such as a website hero or a slide, small faces in a corner of a screen recording, and cheap tests of speech timing.
What is the cheapest vertical option?
Kling AI Avatar v2 Standard at $0.0562 per second, or MiniMax H3 Max Lip Sync at a $0.05 base for very long audio.
Does it accept any aspect ratio?
No. The output is fixed at 768 by 448 at 25 frames per second.
How to do this without code: The talking-photo path only uses models that output a vertical frame. 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: AI Video Models on fal in October 2026: What Is New, What It Costs, and What Is Gone · HappyHorse Video Edit on fal: What It Costs to Replace a Person in a Clip, and the Limits (October 2026)
For developers: the fal endpoint
| Model | fal endpoint |
|---|---|
| SoulX-FlashTalk | fal-ai/flashtalk |
Output fixed at 768×448 at 25 fps. $0.02 per second, read on 18 October 2026.