Flux 3 T2V vs Flux 3 extend

Flux 3 Text to Video is 34 credits, text-to-video, native audio true, 5s through 20s every second, no image required, resolution unset. Flux 3 Extend Video is 82 credits, extend-video, native audio true, the same 5–20s ladder, no image required on the gate. Thirty-four starts a talking clip. Eighty-two continues a keeper. Do not pay 82 to invent a new identity, and do not pay 34 again to lengthen a keeper you already liked.

This is not the routing-philosophy post about extend versus a new generate. That page already named the two failures. This page is the sticker: 34 versus 82, both audio true, both 5–20s. Black Forest Labs’ 4 August 2026 GA note (BFL blog) describes clips up to 20s with native audio. That is BFL’s family story; the bills here are the Versely catalog’s, in credits.

The T2V description: FLUX.3 generates video with native audio from a text prompt, at up to 1080p and 5–20 seconds across eight aspect ratios. The snapshot res field is unset. Do not treat “up to 1080p” in the copy as a filled 1080p token. Do not invent 4K.

The extend description: FLUX.3 continues an existing clip beyond its final frame; source clip must be MP4, under 50 MB and under 15 seconds. That source limit is in the copy. The duration chips on the extend card are still 5–20s. You continue a clip that is under 15s; you pick a hop length from the 5–20s list. Do not feed a 20s file if the description says the source must be under 15s.

Thirty-four credits is the talking generate

Use T2V when you need a new plate. Write the sentence. Write whether anyone speaks. SKU orbits usually want Foley and no voice. Say “no speech.” A 34-credit surprise monologue is still waste, cheaper than 82 of the same mistake.

Length: 5s is the draft. 8s or 10s is a common ship. 20s is the ceiling. The slider going to 20 is not a compliment. Draft short. Promote length on the same prompt only when the short take is already the identity you want — or stop promoting length on T2V and use extend so the face does not resample.

When T2V is the pick:

• There is no keeper yet.
• The idea might be wrong. Resampling is then a feature.
• 5–20s with native audio is the file.
• You have a sentence, not a still that must be frame one. First-last is a different Flux slug at 34 credits.

When T2V is the wrong pick:

• The last frame is good and the camera is still travelling. That is extend at 82.
• You needed a still as frame one. Flux 3 First Last Frame to Video is 34 credits, I2V, native audio true, 5–20s.
• You needed 30s. Flux stops at 20s. Grok Imagine Video starts at 24 credits, 720p, up to 30s, native audio.
• You needed listed 4K. Veo 3.1 starts at 64 credits, 4K, 4 / 6 / 8s, native audio (32 a second with sound on).

Thirty-four credits is the catalog sticker. BFL’s early evals (23 Jul 2026) preferred Flux 3 over several hosts; those are BFL’s results, not a Versely benchmark. The bill here is 34.

Eighty-two credits is the talking hop

Extend keeps motion, room, and audio bed from the keeper. Both rows mark audio true. Unlike Grok Imagine Extend at 13 credits with audio false, Flux’s hop can keep speaking. Write the next beat as continuation, not as a new scene. “Now they are in the car” is a new generate wearing an extend button.

Source constraints from the description: MP4, under 50 MB, under 15 seconds. A 20s T2V hero is not a legal source if that under-15s rule holds. Plan the first pass at 15s or less if you know you will hop. A 5s or 10s keeper is the usual extend input.

Duration chips 5–20s on the hop: you are buying another 5 to 20 seconds of continuation, not a checkbox that “makes it longer by 2s.” The cheapest honest hop on the list is 5s at 82 credits. That is more than two T2V generates. The hop has to be worth not resampling.

When extend is the pick:

• The keeper is a Flux 3 clip (T2V or first-last) whose last frame you would print.
• The next beat is the same shot.
• You want native audio to continue, not a mute stitch.
• The source is MP4, under 50 MB, under 15s.

When extend is the wrong pick:

• The hook is already wrong. Eighty-two credits will grow the wrong take.
• You wanted a different room or lens. That is a new 34-credit generate, then a cut.
• The keeper is Grok, Veo, or Seedance. Stay in family. Grok’s hop is 13 and silent. Veo’s extend is a different slug at 112 credits (224 with sound), 7s, 720p.

Do not extend to repair a line. Change the line on a new generate, or lipsync a locked plate on a lipsync row. Extend continues. It does not rewrite.

34 versus 82 in production order

1. Draft T2V at 5s. Listen. If the identity is wrong, regenerate at 34. Do not extend.
2. If 5s is the ship, stop. You spent 34.
3. If you need up to 20s of the same new idea and there is no keeper yet, you may generate longer on T2V. Know that identity can drift versus a hop off a locked 5s.
4. If the 5s (or 10s) take is the keeper and the camera is still moving, extend at 82 for 5–20s more. Stay under the source size and 15s source cap.

Two T2V takes are 68 credits. One T2V plus one extend is 116. Three T2V takes are 102. Extend is not cheaper than another generate. Extend is how you refuse a new face. Pay 82 for continuity, not for savings.

The AI video generator is the launcher. Captions still belong on a talking Flux file. Captions for silent autoplay is the typesetting. The AI caption generator burns the line. Extend will not caption.

Article 50 still applies to photoreal people. Thirty-four credits does not exempt a generated face. Eighty-two does not either. Label at first exposure for EU audiences since 2 August 2026. The Article 50 disclosure template is the copy.

Sora 2 T2V was 32 credits, native audio, 720p, up to 20s. The Sora API shuts 24 September 2026 (OpenAI Help Centre, September 2026). A leftover Sora 20s talking sentence can land on Flux T2V (34 credits for 5s, 136 for 20s) if you accept the unset resolution field. It does not land on extend unless you already have a Flux keeper under 15s.

UGC that is a premise with no photo yet is T2V at 34. UGC that is a locked take you must lengthen is extend at 82. The brand name Flux will not pick for you.

What 82 is not

82 is not a 20s slider on the T2V card. 82 is not cheaper look-dev. 82 is not Grok’s silent 13-credit hop. 82 is Flux 3 Extend Video: native audio, 5–20s, source under 15s and 50 MB. Thirty-four is the generate. Eighty-two is the continue.

Run the 34-credit start, then the 82-credit hop only if the keeper earned it, in Versely.

FAQ

Does Flux 3 extend stay native audio?

Yes. The published card is audio true, same as Flux 3 T2V. Grok’s extend is the silent hop. Do not mix those flags.

Why is extend 82 if generate is 34?

Extend buys continuity of identity, motion, and bed. Two generates are 68 and may resample the face. Eighty-two is the no-resample bill, not a discount.

Can I extend a 20s Flux take?

The description requires the source clip under 15 seconds. Plan the keeper at 15s or less if you know you will hop.