LTX 2.5 on fal: Open-Weights 22B Video With Multishot and 4K, and a Price Unit Nobody Explains (October 2026)
Lightricks' LTX 2.5 arrived in August 2026 as a 22-billion-parameter open-weights video model with native multishot and 4K HDR. What it offers, why we cannot quote what a clip costs, and why LTX 2.3 at $0.08 per second is still the practical choice.
LTX is the video model family from Lightricks, and the one in it that matters for budget work was LTX 2.3 at $0.08 per second. LTX 2.5, released on 11 August 2026, is the next generation: a 22-billion-parameter model with open weights, which means anyone can download and run it, and which the vendor describes as supporting native multishot generation, an automatic duration choice and 4K HDR output. fal hosts it in Pro and Fast variants for image-to-video, text-to-video and audio-to-video. The sticking point for anyone who needs a budget is a single line on fal's price page.
The price problem
fal lists LTX 2.5 at $0.01 per unit, and does not say what a unit is. It could be a second of output, a frame, a thousand tokens or a GPU-second, and each of those would give a very different bill for a 10-second clip. Compare how other models list: Seedance 2.5 bills in video tokens and says so, MiniMax H3 Max in output seconds at a 768P base, FLUX.3 image in megapixels. LTX 2.5's unit is undefined.
| If a unit is | A 10-second clip costs |
|---|---|
| a second of output | $0.10 |
| a frame (about 24 per second) | about $2.40 |
| a GPU-second | depends on the run |
Those are not predictions, only an illustration of how far the range is. Until one paid run measures a real bill, we treat LTX 2.5 as unquotable and carry it in our catalogue as a candidate rather than an option for users. That is why LTX 2.3, at a clear $0.08 per second, remains the quotable member of the family even though it is a generation behind.
What the vendor says it can do
- Open weights: the model can be self-hosted. For an organisation with GPUs and a privacy requirement that is a different proposition from an API.
- Native multishot: a prompt can describe several shots and the model cuts between them in one generation, which matters for storytelling and for ads.
- Automatic duration: the model picks a length that suits the content.
- 4K HDR: on the vendor's specification; we have not verified it on fal's hosted version.
- Audio-to-video: a clip driven by a supplied audio track.
These come from the vendor's description; we have not generated with the model.
What we would use it for, if the price resolves
Face-less work at volume: product loops, backgrounds, abstract motion, where a cheap open model is attractive. Self-hosting: if a unit cost on your own hardware beats an API, the open weights are the reason to look. Multishot ads, if the shot transitions hold up, which is the thing to test first.
For any clip with a recognisable person, use a model with a record. We have no basis to say how LTX 2.5 treats real faces.
The practical alternatives today
| Need | Quotable choice |
|---|---|
| Cheapest face-less motion | LTX 2.3 at $0.08, or MiniMax H3 Max at a $0.0325 base |
| Controllable face-less scenes | FLUX.3 video at $0.085 |
| 30-second face-less take | Wan 3.0 Prime |
| A person on screen | Kling 3.0 Pro, HappyHorse |
The MiniMax H3 versus LTX 2.3 comparison sets out why cheap and expensive models differ less than the price suggests for work with no face.
How to try it responsibly
Run one 5-second image-to-video clip on the Fast variant, then read the invoice or the usage page before running a second. If a 5-second clip costs a few cents, the unit is likely a second and the model is cheap. If it costs dollars, stop and find out why. Do not scale a workflow on an unpriced endpoint, because nothing in the price list protects you.
Honest status
Everything here is the vendor's specification and fal's listing as read on 18 October 2026. The model is new, the unit is unexplained, and we are not recommending it for production until both are settled.
Common questions
What does open weights mean?
The model can be downloaded and run on your own hardware, which can be cheaper at volume and keeps footage private.
Why can you not quote a price?
fal lists $0.01 per unit and does not define the unit, so the cost of a 10-second clip could differ by orders of magnitude.
Is LTX 2.3 still available?
Yes, at $0.08 per second, and it is formally superseded by 2.5.
Does it support real faces?
We have no basis to say. Use a model with a record for any clip where a recognisable person must stay recognisable.
What is native multishot?
A prompt can describe several shots and the model cuts between them in one generation, according to the vendor.
How to do this without code: Face-less scenes are routed to a generator for you, and the cost is quoted before you confirm. In AutoHustle this is the autopilot on a Reel with no person in it, where face-less scenes are kept from the original footage rather than regenerated (quoted in credits before you confirm the plan). See Faceless content and Recreate a viral video.
Related reading: Kling Native 4K on fal: $0.42 per Second, and When 4K Is Worth Paying For (October 2026) · 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 |
|---|---|
| LTX 2.5 image-to-video Pro | lightricks/ltx-2.5/image-to-video/pro |
| LTX 2.5 image-to-video Fast | lightricks/ltx-2.5/image-to-video/fast |
| LTX 2.5 text-to-video | lightricks/ltx-2.5/text-to-video/* |
| LTX 2.5 audio-to-video | lightricks/ltx-2.5/audio-to-video/* |
Listed at $0.01 per unit; the unit is not documented. Read on 18 October 2026.