Models, Limits, and the August 15 Deadline

Two models, and only two

The exact enum values are ltx-2-3-pro and ltx-2-3-fast. ltx-2-pro and ltx-2-fast are deprecated — hold that thought, it becomes the second half of this lesson.

Which one you can use is not a preference. It is a table:

Fast reaches exactly two endpoints. Everything that edits, extends, reframes, or drives video from an audio track is Pro. video-to-video-hdr sits outside the matrix entirely: it takes no `model` parameter at all — it is model-agnostic.

The judgment this table encodes: Fast is a generation tier, not a production tier. You can build an entire exploration loop on Fast — cheap passes, long-form drafts at 1080p — and then discover that every finishing move you want is Pro-only. Decide early which of your shots will need retake, extend, or reframe, because those shots live in Pro's world and Pro's world caps at ten seconds.

The duration table

This is the fact people get wrong most often, so learn it as a table rather than as a sentence:

Maximum resolution is 4K — 3840×2160 or 2160×3840 — on both models.

Read down the right-hand column. There is exactly one row in the entire matrix that gives you 20 seconds: Fast, at 1080p, at 24 or 25 fps. Every other combination in existence caps at 10. Push Fast to 48 fps and you lose half your ceiling. Push Fast to 1440p and you lose half your ceiling. Choose Pro for any reason at all and you are at 10 seconds, at every resolution, at every frame rate.

What a 10-second ceiling means for how you work

It means you think in shots, not scenes. This is not a limitation to route around; it is the unit of work the tool gives you, and it happens to match how film has always been assembled.

Some arithmetic. A 60-second piece, generated on Pro:

• At 10s per shot: 6 generations minimum, and that is with zero cuts inside a shot, which no edit ever survives.
• Realistically 12 to 15 shots averaging 4 to 5 seconds, which is a normal cutting rhythm anyway.

The same 60 seconds on Fast at 1080p, 24 fps, at 20s a shot, is three generations — but you have then given up retake, extend, reframe, and audio-to-video for those shots. That is usually the wrong trade for finals and the right trade for previz.

extend exists to add frames at the head or tail using a context window, and its billing is capped at 505 billed frames total — a ceiling, not a licence to grow a clip indefinitely. Treat extend as continuity repair at a shot boundary, not as a duration workaround.

The August 15 deadline

Verbatim from the deprecation notice:

Today is August 5, 2026. You have ten days.

Understand the specific danger in that quote. Right now, code sending ltx-2-pro works. It returns video. Your tests pass, your dashboards are green, your creators are happily producing. Deprecation was not a breakage — it was a silent redirect, with legacy billing attached. Nothing in your normal operation will tell you that you are on the doomed path.

The notice says removed requests "return an error" and does not name the error type. Do not guess at it, and specifically do not write a retry wrapper that swallows it. A hard failure you can see on August 16 is worth more than a quiet retry loop burning your rate limit against a model that no longer exists.

The lab below is the audit. It takes about twenty minutes. Do it now rather than on the fifteenth.