MiniMax H3, Sold by the Second

Open-weight video models are thrilling — and a little chaotic — when you’re wiring them into a real production loop. On September 17, 2026, Munich/Paris efficiency startup Pruna AI shipped P-Video-2-Pro: a commercial API endpoint built on MiniMax H3 (open weights since August 3) that packages text-to-video and first/last-frame image conditioning into something pipelines can actually call, meter, and iterate on.

This isn’t another “here’s H3 again” rehash. ArtRealmAI already covered the open MiniMax H3 story. Today’s scoop is the speed/API packaging: resolutions, modes, upsampler knobs, per-second pricing, and the honest tradeoffs versus full H3.

What Ships in the Box

P-Video-2-Pro generates 5–15 second clips at 24 fps with generated audio baked in. Audio input isn’t exposed — you don’t feed a soundtrack; the model invents one. Resolutions are 480p and 768p. Aspect ratios cover the usual creator grid: 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, and 1:1 (reference images set the canvas when you pass frames).

Conditioning is refreshingly practical for storyboards and product shots:

  • Text-to-video from a prompt alone
  • First-frame image conditioning
  • Last-frame image conditioning (or both ends for a guided arc)

There’s a clean split between generation recipe and prompt expansion:

  • mode: speed (default) vs quality
  • prompt_upsampler: off / turbo (default) / max — independent of mode, and it doesn’t change the price

Vendor latency claims (speed mode, 5s clips): about ~2s for 480p and ~4.3s for 768p. Treat those as Pruna’s numbers, not an independent lab bench — but if they hold in your region, that’s the kind of turnaround that makes prompt-sweeping feel like editing, not waiting.

Pricing That Matches Iteration

Billing is per finished second of the returned video (not the request duration you hoped for), so the invoice matches what you actually got. List prices:

ConfigPrice / finished second

480p · speed

$0.02

480p · quality

$0.04

768p · speed

$0.035

768p · quality

$0.075

50% off through September 24, 2026. Rate limit: 250 requests per minute. Omitted mode bills as speed. Upsampler choice is free of charge — nice when you’re still hunting the right prose.

For a five-second 480p speed preview that’s $0.10 list / $0.05 on the promo. Cheap enough to burn a dozen angles before you spend for a 768p quality hero.

Narrower Than Full H3 — On Purpose

Full MiniMax H3 is a kitchen-sink open model. P-Video-2-Pro is deliberately narrower: no 2K, no audio/video reference inputs. That sounds like a downgrade until you remember what most app backends need — predictable endpoints, billed seconds, and a mode dial for “draft vs ship.” Pruna’s pitch is pipeline iteration: generate, judge, tighten the prompt, regenerate, without babysitting a local Comfy graph every time.

Pruna itself is the efficiency shop behind an open-source optimization framework (~1.3k GitHub stars) and a $6.5M EQT-led seed (November 2024). Selling fast H3 inference fits the brand: squeeze quality into latency and cost budgets creators can live with.

How to Call It (Sketch)

Docs show the familiar Pruna predictions shape — Model: p-video-2-pro, apikey header, optional Try-Sync: true for synchronous short jobs. A minimal text-to-video body looks like:

curl -X POST 'https://api.pruna.ai/v1/predictions' \
  -H 'Content-Type: application/json' \
  -H 'apikey: YOUR_API_KEY' \
  -H 'Model: p-video-2-pro' \
  -H 'Try-Sync: true' \
  -d '{
    "input": {
      "prompt": "A sports car drifting through a neon-lit city at night, cinematic aerial shot",
      "duration": 5,
      "resolution": "480p",
      "mode": "speed",
      "aspect_ratio": "16:9"
    }
  }'

For image-to-video, upload via /v1/files, then pass the file URI as image and/or last_frame_image. Seed is optional for reproducibility.

Who This Is For

If you’re prototyping ads, game cinematics, product loops, or social verticals and you want H3-class motion with an invoice line item, this is the packaging story to watch. Keep open-weight H3 for research and self-host experiments; reach for P-Video-2-Pro when your app needs a stable commercial path with speed/quality modes and promo pricing through Sep 24.

Original Source

Docs: p-video-2-pro API guide · Product: P-Video-2-Pro · Context: RuntimeWire launch write-up

Conclusion

P-Video-2-Pro is MiniMax H3 dressed for commerce: 5–15s clips, generated audio, first/last-frame conditioning, speed-vs-quality knobs, and per-second pricing with a short 50% off window. Narrower than the full open model — and that’s the point. If your pipeline lives or dies on fast API iteration, this is the H3 dial worth turning before the promo ends.

—Aurelia ♡