Pruna P-Video and P-Image Models: Draft-Mode Speed and Pricing (2026)
Last updated August 7, 2026
Pruna AI's P-models are the speed and cost floor of the 2026 model landscape: P-Video (Feb 2026) drafts a 5s 720p clip in 2.5 seconds and charges $0.02-0.04/s at full quality, with Avatar, Animate, and Replace variants; P-Image renders 1024px images in 0.6s for $0.005, with an Ideogram typography collaboration. Quality sits below frontier by design. All run inside the invideo agent.
Updated August 2026
Pruna AI's P-models are built around one bet: most generation work is iteration, and iteration should be nearly free. P-Video (February 2026) generates a draft of a 5-second 720p clip in 2.5 seconds and charges $0.02–0.04 per second for full-quality output; P-Image returns a 1024px image in 0.6 seconds for $0.005. These aren't frontier-quality models and Pruna doesn't pretend otherwise — they are the speed and cost floor of the current model market, from a company whose whole business is making AI models "faster, smaller, cheaper, greener." The P-Video and P-Image families both run inside the invideo agent.
What is P-Video's draft mode, and why does it matter?
P-Video (launched February 26, 2026) is a text-, image-, and audio-input video model with audio output, supporting 720p and 1080p at up to 15 seconds. Its signature feature is a two-stage workflow: draft mode produces a 5-second 720p clip in 2.5 seconds (4.9 seconds at 1080p), and full quality takes 10 seconds for the same 720p clip (18 seconds at 1080p).
The economics are the point. Prompting video is trial-and-error, and on premium models every failed attempt costs money and waiting. A 2.5-second draft loop lets you burn through ten prompt variations in under a minute, then spend full-quality pricing — $0.02/s at 720p, $0.04/s at 1080p — only on the winner. Pruna claims the "best speed-to-quality-to-cost ratio on the market" based on a VBench comparison; note that this is Pruna's own benchmarking, not an independent evaluation.
What's in the P-Video family?
| Model | What it does | Official price (Aug 2026) |
|---|---|---|
| P-Video | Text/image/audio-to-video with audio, 720p/1080p, up to 15s, draft mode | $0.02/s (720p), $0.04/s (1080p) |
| P-Video Avatar | Portrait-to-talking-head — makes a still speak from a script or audio | $0.02/s |
| P-Video Animate | Animates one still using a driver clip, keeping the image's atmosphere | $0.025/s |
| P-Video Replace | Swaps a character into existing footage, keeping the video's atmosphere | $0.03/s |
The spread between the cheapest and priciest variant is a single cent per second — pick by task shape, not by price. Pruna publishes prices and one-line descriptions for the specialist variants but little deep documentation, so treat the flagship P-Video specs as the family reference.
What's in the P-Image family?
- P-Image: real-time-class text-to-image — 0.6s inference, 1024px, $0.005 per image.
- P-Image Edit: instruction-based editing — 0.9s, $0.010 per image; both live on third-party creative platforms since December 2025.
- P-Image Ideogram: a confirmed collaboration — an Ideogram-built, Pruna-optimized model with strong typography, up to 2048px, four "thinking levels" (Very Low to High) trading latency (~1.3–5.0s) against quality, at roughly $0.003–0.015 per image.
- P-Image Upscale: fast upscaling — 0.9s at the base tier, $0.005 per 4MP run; in invideo's model picker it's exposed with output targets from 1MP up to 128MP.
Who is Pruna AI?
Pruna is a Franco-German model-optimization startup (Paris and Munich, founded 2023) that raised a $6.5M seed in November 2024 and open-sourced its optimization engine in March 2025 — a toolkit applying compilation, caching, quantization, and distillation to arbitrary models. Through 2025 it optimized other labs' models — its "juiced" Wan 2.2 endpoints (mid-2025) were billed as the fastest, cheapest way to run Wan. The P-models, arriving from December 2025 onward, are Pruna's own branded inference line: the same optimization stack, productized. One caveat, stated plainly: Pruna does not disclose which base models the P-line is built on, as of August 2026.
Strengths and honest limitations
Strengths. The latency numbers are the story: sub-second images and 2.5-second video drafts change how iteration feels, and the per-unit prices sit at or near the market floor as of August 2026. The open-source engine lends the speed claims engineering credibility, and the Ideogram collaboration adds a genuine typography specialist to an otherwise generalist line.
Limitations. The quality ceiling sits below frontier video and image models — Pruna positions the P-line on speed and cost, and its favorable comparisons are self-published (VBench-based and its own InferBench). Base-model provenance is undisclosed. Documentation on the specialist variants is thin. If you need one hero shot at maximum fidelity, this is the wrong family; if you need two hundred usable shots by Friday, it might be the right one.
What should you use P-models for?
- High-volume drafting — exploring prompts, compositions, and motion ideas in an AI video generator workflow before committing a premium model to the final render.
- Storyboards and animatics — P-Image's 0.6-second turnaround makes boarding a 30-shot sequence in a storyboard session practical rather than painful.
- Iteration-heavy image work — thumbnails, variants, and A/B explorations in an AI image generator loop, switching to P-Image Ideogram whenever legible text matters.
Workflow tips
Treat drafts as disposable: iterate wording, composition, and camera movement in draft mode, and re-render only winners at full quality. Keep prompts concrete and short. Route text-heavy frames to P-Image Ideogram, raising its thinking level only for complex layouts. And use P-Video drafts as previsualization even when a premium model renders the final — a cheaply refined prompt usually transfers.
Quick answers on the P-models
How much does Pruna's P-Video cost?
$0.02 per second at 720p and $0.04 per second at 1080p, per Pruna's official announcement (February 2026). P-Video Avatar is $0.02/s and P-Video Animate $0.025/s.
What base models are the P-models built on?
Pruna doesn't say. The company previously optimized third-party models (such as Wan 2.2), but the P-line's underlying architecture is undisclosed as of August 2026.
Is P-Video good quality?
It targets the best speed-to-cost ratio, not the quality ceiling — Pruna's own VBench-based comparison backs the ratio claim, but frontier models beat it on pure fidelity.
What is P-Image Ideogram?
A collaboration model: built by Ideogram, optimized by Pruna, specializing in typography and text rendering at up to 2048px, with four selectable "thinking levels."
Are Pruna's models open source?
The optimization engine is open source (March 2025); the P-model weights themselves are not.
Where can you run the P-models?
The whole family — P-Video, P-Video Avatar, P-Video Animate, and the P-Image line — is available through the invideo agent, where the practical pattern is to draft with a P-model and re-render keepers on a heavier engine without switching tools. They're listed on the AI models index; as the AI video platform that gives serious creatives every major model in one place, invideo treats the P-line as its fast lane rather than its finish line.
Version history: first published August 2026, covering P-Image and P-Image Edit (December 2025), P-Video and the Avatar/Animate/Replace variants (February 2026), and the P-Image Ideogram collaboration.