Pruna offers MiniMax H3 video models with output up to 4K

Since Unicorn’s last report, which covered Pika’s API documenting MiniMax H3 at 768P or 2K, a second developer platform lists H3 in its catalogue. Pruna’s developer portal now offers three video models it says are “built on MiniMax H3”, with output it prices from 480p up to 4K. The three cover frame-guided generation, reference-guided generation and talking avatars. The evidence is Pruna’s own model list. None of the sources provide independent quality tests.

What Pruna lists on top of MiniMax H3

Pruna’s “Available Models” page names three products built on H3, and gives each its own pricing and rate limit:

• p-video-2-pro: described as “Premium video generation from text and first/last-frame images, built on Minimax H3”, priced at $0.01–$0.155 per second “by resolution (480p to 4k) and mode”, with a rate limit of 250 requests per minute.
• p-video-2-pro-reference: “Video from a prompt with image, video, and audio references, built on MiniMax H3”, priced at $0.015–$0.115 per second of output by resolution and mode, “plus reference media”, also at 250 per minute.
• p-video-2-pro-avatar: “Talking video from a portrait and audio or a script, built on MiniMax H3”, priced at $0.02–$0.12 per second of output from 480p to 4K, at a lower 50 per minute.

“First/last-frame” generation means the developer supplies the opening and closing images and the model generates the motion between them. “Reference” generation means the model takes example images, clips or audio as guidance for characters, motion or timing, rather than relying only on a text prompt.

The same page also lists a separate p-video-upscale model that “Upscale[s] a source video to 720p, 1080p, 2k or 4k, keeping its aspect ratio and audio”, at $0.01–$0.16 per second of output. Pruna does not describe the upscaler as part of the H3 family.

Pruna model	Inputs (per Pruna)	Price per second (per Pruna)	Rate limit
p-video-2-pro	Text, first/last-frame images	$0.01–$0.155, 480p to 4K	250/min
p-video-2-pro-reference	Prompt plus image, video and audio references	$0.015–$0.115 output, plus reference media	250/min
p-video-2-pro-avatar	Portrait plus audio or script	$0.02–$0.12 output, 480p to 4K	50/min

All figures in the table come from Pruna’s model list.

How this compares with Pika’s H3 offering

Unicorn’s earlier article reported that Pika’s API documents H3 for text-, image- and reference-to-video at 768P or 2K, for clips of 4 to 15 seconds. Pruna’s list differs in two ways. It goes further on resolution, naming tiers from 480p to 4K. It also splits H3 into separate products by task, including a talking-avatar model, rather than one endpoint with several modes.

Pruna’s model list does not state a maximum clip length for its H3-based models, so the sources do not allow a direct comparison on duration.

APIDot, a third API reseller, publishes a set of guides on the same model under the name “Hailuo 03 (MiniMax H3)”. Its review covers “native 2K video, 5–15 second clips, native audio, multimodal references, first-and-last-frame control, API access, and current pricing”. Its pricing guide refers to “21 credits per second” on APIDot’s own credit system. APIDot also publishes guides on image, video and audio references and on prompting H3 for “5–15 second native 2K videos”.

A resolution question builders should check

APIDot describes H3’s native output as 2K. Pruna lists its H3-based models with output up to 4K. Pruna’s page does not say whether its 4K tier is generated natively or produced by upscaling, and it sells a separate upscaler that reaches 4K. The two descriptions may both be accurate, but the sources do not settle it. Teams that need true 4K detail should test outputs at each tier rather than rely on the label.

Access and pricing as the sources state them

All the access described in the sources is through hosted APIs. Pruna lists the models on its developer portal with per-second pricing and per-minute rate limits. APIDot’s guides cover “API access” to H3. Pika’s own API, covered in Unicorn’s previous article, documents the model as well.

Pruna’s prices are the provider’s own and vary widely within each model. The ranges depend on resolution and, for two of the three, on a “mode” setting. For the reference model, the cost of the reference media comes on top of the output price, and Pruna’s list does not break that cost down. Budgeting for reference-heavy workflows therefore needs a test run, not just the headline range.

The rate limits matter for production planning. The two general H3 models allow 250 requests per minute, while the avatar model allows 50. That is the same 50-per-minute limit Pruna lists for its other avatar, animation and replacement video models.

Where Pika fits in the video tool market

Third-party comparison guides published this year position Pika as a fast, lower-cost tool for social video, separate from its developer API. One guide describes Pika as “designed for simplicity and rapid content creation, especially for social media and quick marketing campaigns”, and contrasts it with Runway’s node-based workflows and API access. The same guide notes Pika’s “Pika Agent” for creative collaboration and “Pika MCP” for enhancing AI agents. MCP, the Model Context Protocol, is a standard way for AI agents to call outside tools.

FluxNote’s comparison says Pika’s Standard plan costs $8 per month, against $12 per month billed annually for Runway’s Standard plan. It cites Q1 2026 tests in which Pika produced a 3-second clip in 20–40 seconds, while a comparable 5-second Runway Gen-3 clip took 45–90 seconds. A second FluxNote guide says Pika’s maximum clip length is “often around 4-10 seconds depending on the feature used” and that its photorealism ceiling is lower than Runway’s.

These consumer guides do not mention MiniMax H3, and they disagree on details. AICentralResources gives Pika’s free tier as 80 monthly credits in one place. Elsewhere on the same page it gives 150 credits per month, and also 250 initial credits plus 30 daily credits. Their clip-length figures describe Pika’s own tools and may not apply to H3 jobs run through Pika’s API.

What it means for builders

H3 is now available from more than one hosted provider, which gives developers options on price, rate limits and packaging. Pruna separates frame control, reference control and avatars into distinct endpoints, each with its own price and limit. This can simplify routing in an application: a product that only needs talking-head video can call the avatar model directly.

The limits are just as clear:

• No independent quality evidence. The sources contain no third-party benchmark or leaderboard result for H3. Descriptions such as “Premium” are Pruna’s own.
• Naming varies. APIDot calls the model “Hailuo 03 (MiniMax H3)”, while Pruna writes both “Minimax H3” and “MiniMax H3”. Developers searching documentation should check both names.
• Feature parity is unconfirmed. APIDot lists native audio among H3’s features. Pruna’s H3 model descriptions mention audio as a reference input and as an avatar driver, but do not say whether generated clips include native sound.
• Undisclosed modifications. Pruna says its models are “built on” H3. It does not say whether that means direct access to the base model or a modified version.

What to watch next

The open questions are concrete. Builders will want to know whether Pruna’s 4K tier is native or upscaled, and what clip lengths its H3 models support. They will also want to see whether any third-party evaluation of H3 appears to back the providers’ descriptions.

This thread was opened to track a reported Pika variant of H3. None of the sources reviewed for this article confirm an open-weights release of H3, or of any H3 variant, from Pika or MiniMax. Unicorn will report one only when a primary announcement or model card is available. Until then, every route to H3 described in the sources is a hosted, paid API.