One table, three different tests

Start with what LTX got right. Most vendor benchmarks hide their conditions. This one publishes them: the task, the frame rate, the resolutions, the hardware, and the admission that hosted numbers include queue time while the self-hosted number is a steady-state run. A reader who reaches the methodology section can reconstruct exactly what was measured. That is more disclosure than this category usually offers.

The problem is where the disclosure sits. The headline result, repeated in the summary, is that LTX-2.5 made a 10-second clip in 6.8 seconds while Kling 3.0 Pro took 398. Read as one race, the published times stand in a ratio of about 58 to 1. It was not one race. LTX-2.5 ran on LTX's own hardware, two GB200 GPUs, warmed up and waiting for nobody. Kling ran through fal.run, a shared provider, and its 398 seconds include however long the job spent in line behind other customers. One number is inference speed. The other is inference speed plus a queue of unknown length. LTX's own text concedes the two do not compare directly, then builds its iteration arithmetic on them anyway: 40 variations, it says, cost about 4.4 hours on the slowest model and under 16 minutes on LTX. Check which LTX number that uses. Forty sequential requests at 398 seconds each come to roughly 4.4 hours; at 23.7 seconds each, roughly 15.8 minutes — so the "under 16 minutes" figure belongs to the API time, not the 6.8-second self-hosted one. Either way, the arithmetic is an extrapolation from single reported times: forty requests run one after another at unchanged latency, with no concurrency and no queue behaviour under load. It illustrates the gap; it does not measure a production workload.

One more correction belongs here, about the date. The 6.8 seconds is not a new record set for this post. The figure was already attached to LTX-2.5 in its launch coverage in August. What is new on 19 October is the comparison wrapped around it.

The comparison that survives scrutiny

Strip the self-hosted row out and a narrower claim remains. Through its API, LTX-2.5 produced a 10-second clip at 1080p in 23.7 seconds. For hosted buyers that is the more relevant row in the table: it is the service they would actually call.

But it is not a controlled comparison either, and it should not be read as one. The 23.7 seconds is an internal LTX measurement. LTX does not say whether any queue time sits inside it, does not specify the infrastructure behind it, and does not say which API variant, Fast or Pro, produced it. The rivals, by contrast, were timed end to end on fal.run with queues included. So LTX reports the shortest hosted generation time among the services listed in its comparison, and the measurements still do not establish a controlled ranking of hosted performance. The competitors and the conditions were both chosen by LTX.

The ratios change with the yardstick, which is the point. Against Kling's 398 seconds, substituting the API time changes the arithmetic from about 58× to about 17×; against the fastest hosted rival, Omni Flash at 52 seconds, it is about 2.2×. Those are ratios between published observations taken under different conditions, not demonstrated speed advantages — and because the queue share inside the rival times is unknown, removing it could materially change every one of them.

The quality chart needs the same discount, and a clearer separation. It comes from a different test: 98 text-to-video prompts, graded by automated scoring, which LTX labels preliminary. LTX-2.5 Pro's 0.28 artifacts per clip leads the field shown, and LTX notes in the text that Seedance 2.0 scored 0.60 and is not shown on the chart. A vendor leaving a mid-pack rival off its own chart is a small thing, and exactly the kind of small thing that tells you how to read the rest of the page. Either way, a text-to-video artifact count does not validate the image-to-video outputs in the speed test, and it says nothing about overall visual quality, prompt adherence or what users would prefer.

What the speed is actually attached to

The part of the post with the longest shelf life is not the ranking. It is the combination underneath it: 22B open weights you can download, fine-tune and run on your own hardware, with a managed API at $0.09 per second at 720p if you would rather not. No closed rival in the table offers that combination, and LTX is right that speed plus ownership is a different proposition from speed alone.

The licence has a name and conditions, and both belong next to that claim. The weights are released under the LTX Community License. On LTX's own pages, organisations under $10M in annual revenue may use them commercially at no licence fee, subject to the agreement's conditions; organisations above that line need a paid commercial agreement. That is open weights with terms attached — a real and useful offer, but not unrestricted open source, and the agreement is worth reading before a pipeline depends on it.

Two more conditions. The full model lists 80GB or more of VRAM; the lighter distilled and FP8 variants list 32GB. Neither is the machine the 6.8-second run used, and neither will reproduce it. And every figure in the post — speed, quality and the rest — was published by the company selling the model. None of it has been run by an independent lab in the form presented here.

Run your own prompts, on the tier you would actually pay for, and time the queue separately from the generation. LTX's post, read in full, almost tells you to do exactly that. Read only its table, and you would think the race was already over.

Sources

  1. [1] LTX (official), "Fastest AI Video Generation Models (2026): LTX-2.5, Veo, Kling, Seedance Compared", published 19 October 2026 08:08 UTC — read in full, including the methodology sectionRead source
  2. [2] LTX (official), LTX-2.5 model page — speed FAQ (2× GB200, 720p, 6.8s), API pricing per second by tier and resolution, licensing summaryRead source
  3. [3] LTX (official), LTX AI Video Model Commercial License page — the LTX Community License: commercial use free under $10M annual revenue, commercial licence required aboveRead source
  4. [4] Secondary corroboration for the 11 August 2026 LTX-2.5 release date: aivideos.eu, "LTX 2.5 Fast vs Pro vs Audio to Video", and Wikipedia, "LTX (world model)" — release date only; no benchmark figures were taken from secondary sourcesRead source