Lightricks' LTX-2.5 Undercuts Veo on Price, Not on Openness

Lightricks released LTX-2.5 on Aug. 11, a 22-billion-parameter open-weights video and audio model that generates a 10-second clip in 6.8 seconds on two Nvidia GB200 chips, co-founder and chief executive Zeev Farbman said.

The model's weights are posted on Hugging Face under the Lightricks organisation, and it shipped with day-one support in ComfyUI through a launch partnership with that project's co-founder and chief executive, Yoland Yan, according to VentureBeat, which covered the release the same day. A managed API is also available for teams that do not want to run the weights themselves.

What 22 billion parameters buys: multishot scenes and 4K audio-video

LTX-2.5 handles text-to-video, image-to-video, video-to-video, text-to-audio and audio-to-video generation from a single 22-billion-parameter model, according to its Hugging Face model card. It introduces native multishot scenes that keep a character or style consistent across cuts, auto duration selection, and output up to 4K with HDR. The architecture pairs the generation model with a Gemma 4 12-billion-parameter text encoder, and clips are capped at 121 frames by default, with frame counts required to satisfy the formula frames modulo 8 equals 1.

The price undercuts Google's Veo, but the licence is not open source

Lightricks prices the managed API at $0.09 a second for its Fast tier and $0.12 a second for Pro, putting a 10-second 720p clip at 90 cents to $1.20. That is cheaper than Google's Veo 3.1 Fast, listed at $1.00 a clip, and further below FLUX 3 Video at $1.70 and the premium Veo 3.1 tier at $4.00, VentureBeat reported.

Model	Reported price per 10-second clip
LTX-2.5 Fast	$0.90
Google Veo 3.1 Lite	$0.50
Google Veo 3.1 Fast	$1.00
FLUX 3 Video	$1.70
Google Veo 3.1 (premium)	$4.00

The weights themselves are free only for organisations under $10 million in annual recurring revenue under what Lightricks calls the LTX-2.x Community License, according to the model's Hugging Face page. Above that threshold, a company needs a separate paid Commercial Use Agreement, and transferring a fine-tuned version of the model can itself trigger a licensing requirement. VentureBeat's report said plainly that this makes LTX-2.5 open-weights rather than open-source by the term's stricter definition, since usage rights are conditioned on revenue rather than granted unconditionally.

The speed claim comes from Lightricks' own hardware, and the quality score is unverified

The 6.8-second generation time was measured on two GB200 superchips, hardware few developers have access to; through LTX's own managed API, the same 10-second clip at 1080p took 23.7 seconds, VentureBeat reported. Lightricks also commissioned a human-preference test showing a 67 percent win rate against unnamed rivals, which VentureBeat described as preliminary and not independently verified. Artificial Analysis, the leaderboard that scored Qwen3.8-27B at 52 on its Intelligence Index in August, had not added LTX-2.5 as of VentureBeat's Aug. 11 report. The licence also prohibits military and weapons applications, the outlet said.

Whether Artificial Analysis eventually scores LTX-2.5 against Veo and the rest of the field named in Lightricks' own pricing comparison is the test that would move the 67 percent figure from a vendor claim to an independent one.

Lightricks is not the first vendor this month to have a capability claim questioned before independent scoring catches up: Cactus Needle 3's DeepSeek-level tool-calling claim ran into the same gap, and DeepSeek's own V4 Flash pricing shift shows how fast an efficiency pitch can be revised once a smaller model matches a bigger one.