Pruna P Video 2 Is Available Now: Use Cases, API Guide and Exact Costs

Pruna's P Video 2 went live today, and the headline is not the resolution or the length. It is that the model generates sound in the same pass as the picture, at $0.025 per second of 720p output. That is the cheapest per-second video rate on Segmind right now by a factor of at least two, and against most of the models people actually compare it to, a factor of five or six.

I spent $2.53 on it before writing any of this: 15 generations on p-video-2 plus one still on Nano Banana 2 for an image-to-video test. Every fire is accounted for below, with the billed cost taken from the x-cost response header rather than estimated. Two things came out of that which are not in any documentation, and one of them will save you money on the first day: the seed parameter does not work under the default settings, and there is a one-line fix.

What P Video 2 actually is

One endpoint, three ways in. Send a prompt and you get text-to-video. Add an image and it animates that frame. Add an audio track and it generates video to match the audio, which is the shortest path from a voiceover you already own to a presenter clip. The model picks the task from which fields you send, so there is no mode switch to get wrong.

Output tops out at 1080p, 24 or 48 fps, up to 20 seconds, in seven aspect ratios. Every clip comes back as a 48 kHz stereo AAC track married to H.264 video in a single MP4, and the response is synchronous: you POST and the file arrives in the body. No polling, no job IDs, no queue to babysit.

Pruna positions it as the quality-focused successor to P Video rather than a straight speed upgrade, and their partner kit is explicit that the gain is in native-speech lip sync, close-up sharpness and identity lock. It also tells you when not to reach for it: use P Video Edit to modify footage you already have, and their P Video Avatar for single-presenter talking heads with no music bed.

What it costs, measured to the cent

Pricing is per second of output, published as four rates and nothing else:

I fired every meaningful configuration and compared the x-cost header against that table. It matched every time, to the cent, with no per-call fee and no minimum:

The rule is that you pay for the whole seconds the model targeted, and the file you get back lands within one frame of that. Three consequences worth knowing before you budget:

48 fps is free. A 5-second 1080p clip cost $0.250 at 24 fps and $0.250 at 48 fps. I checked that the extra frames are real rather than duplicated: of 241 frames, 240 survived a duplicate-frame filter.

Silence costs the same. Setting save_audio: false strips the audio stream out of the container entirely, which is what you want for Meta autoplay placements, but it still billed $0.125. You are not paying for the audio, so there is no saving in turning it off.

Five use cases, with what each one cost

1. Dialogue to camera, which is the actual headline

The clearest gap between P Video 2 and its predecessor is generated speech. Put the line in quotation marks and the model treats it as dialogue rather than description.

Identity holds across the clip, the close-up is genuinely sharp, and the ambience I asked for is present rather than generic room tone.

2. Product stills into listing video

Retail is where image-to-video earns its keep, because the product has to stay exactly itself. I generated a hero still on Nano Banana 2 for $0.08, then animated it.

The dripper survives intact: same glaze, same silhouette, same plinth. Two mechanical notes. I deliberately sent aspect_ratio: "9:16" alongside the image and it was ignored, exactly as documented, with the frame taken from the image instead.

3. Vertical social

Worth stating plainly because it is not universal: the portrait clip came back as genuine portrait pixels, not a landscape render rotated inside a tall container. I have seen other video models get that wrong and bill for it anyway.

4. Long-form, at the 20-second ceiling

Twenty seconds of coherent single-take footage for fifty cents is the number I would put in front of a finance team. Identity and set dressing held from the first frame to the last, with no visible drift or degradation late in the clip, and the spoken line landed where the prompt put it. Pruna is upfront that multi-scene storytelling is not what the model is for, and I would agree: this works because it is one continuous shot.

5. Draft mode, for exploring prompts

At $0.015 per second, ten 5-second prompt variants cost 75 cents, so exploration stops being a budget conversation. One expectation to reset though: draft mode is 40% cheaper but it was not much faster through the API. Draft returned in 23.5s against a 15.2s to 34.0s spread for full quality at the same length. Pruna's internal figures put draft at roughly twice the model speed, and I have no reason to doubt that, but the model is not the slow part of the round trip. Treat draft as a discount, not a fast lane.

The API, in one call

It is a plain synchronous POST that returns the MP4 in the body:

import requests

r = requests.post(
    "https://api.segmind.com/v1/p-video-2",
    headers={"x-api-key": API_KEY},
    json={
        "prompt": (
            "Close-up of a barista in a sunlit coffee shop, steam curling off the cup. "
            "She looks into the camera and says warmly: \"We roast every batch on Tuesday "
            "mornings.\" The espresso machine hisses behind her, low cafe chatter, no music. "
            "Static camera, shallow depth of field."
        ),
        "resolution": "720p",
        "duration": 5,
        "aspect_ratio": "16:9",
        "fps": 48,
        "seed": 12345,
        "prompt_upsampling": False,
    },
    timeout=300,
)
r.raise_for_status()
open("out.mp4", "wb").write(r.content)
print("billed", r.headers.get("x-cost"))

Four notes that will save you time. Read x-cost on every response, because it is the only place per-call spend is reported. Set your client timeout well above 60 seconds: my slowest fire took 45.0s and the spread is wide. Write the sound into the prompt, since ambience and effects are generated rather than sampled, so naming them is what makes them appear. And validation errors are free and instant: sending duration: 25, duration: 0 or resolution: "4k" each returned a clear 400 in under a second and billed $0.00, so probing the edges of the parameter space costs nothing.

Where the price actually sits

Pruna's kit ships a cost comparison, and I checked its competitor column against Segmind's own published prices rather than repeating it. Per second of 720p output:

Rates with a tilde are computed from published per-million-token rates rather than a flat per-second price, so treat them as close approximations. The rest are arithmetic straight off Segmind's published duration tables.

What I would not claim yet

Pruna is unusually candid about the limits, and their list matches the shape of the model: it is not built for extreme camera moves or multi-scene storytelling, there is no native 4K, speaker separation degrades above two voices, and sound effects are the weakest part of the audio. I did not stress-test those, so I am passing them along as their assessment rather than mine.

Pruna's gaming guide warns that 1080p at 48 fps can return black frames on some runs. It did not happen to me: average luma held steady across the whole clip. One clean run is not a refutation of an intermittent fault, so if you batch at that combination, check your output.

FAQ

How much does Pruna P Video 2 cost?

$0.025 per second of 720p output and $0.050 at 1080p, halving to $0.015 and $0.030 in draft mode. A 5-second 720p clip is $0.125 and a 20-second one is $0.500. I verified every rate against the x-cost header across 15 generations.

Does Pruna P Video 2 really generate audio?

Yes, in the same pass as the video. Every clip returned a 48 kHz stereo AAC track. Set save_audio: false for a silent file, though it costs the same. If you supply your own audio track, the model returns that track unchanged rather than generating new sound.

Why does the seed not reproduce my output?

Because prompt_upsampling defaults to true and rewrites your prompt unseeded on every call. Send prompt_upsampling: false with a fixed integer seed and reruns come back byte-identical.

How long can a clip be?

One to 20 seconds. Omit duration and the model picks a length from the prompt, billing the length it delivers. Supply an audio track and the audio sets the length, overriding whatever duration you sent.

What resolution and frame rate do I get?

720p returns 1280x704 and 1080p returns 1920x1088, at 24 or 48 fps for the same price. With an input image the frame shape comes from the image instead: my 720p image-to-video run returned 1152x768.

Is the Pruna P Video 2 API hard to integrate?

It is one synchronous POST to https://api.segmind.com/v1/p-video-2 that returns the MP4 in the response body. No polling, no job IDs. Budget a client timeout of 300 seconds and read x-cost for the billed amount.

When should I use P Video Edit or P Video Avatar instead?

Use P Video Edit to alter footage you already have, since P Video 2 only generates. Use P Video Avatar for a single presenter reading a script with no music or effects underneath.

Worth your first ten dollars

P Video 2 is the cheapest way to get finished video with sound out of an API right now, and the gap to the next option is not marginal. Ten dollars buys 400 seconds of 720p or 666 seconds in draft. The one thing I would do before you scale anything: send prompt_upsampling: false and a fixed seed, so that when a take works you can actually get it back. Try P Video 2 on Segmind.