Pika's API offers MiniMax H3 video: 2K clips up to 15 seconds

Pika’s developer API now documents MiniMax H3, which it calls MiniMax’s newest video model, for generating 2K text-to-video clips. The model is exposed as text-to-video, image-to-video and reference-to-video endpoints, with output at 768P or 2K. Pika prices it per second of output and charges only for successful runs.

For builders, this means a documented REST route to MiniMax’s current video model through a third-party platform. The listing gives a clear picture of the parameters, the job lifecycle and the costs. It does not include independent quality evidence, so the evidence here is the provider’s own documentation.

What Pika’s API exposes for MiniMax H3

The text-to-video endpoint is minimax/h3/text-to-video. A request takes a required text prompt plus optional settings:

• duration in whole seconds
• resolution
• aspect ratio
• a seed for repeatable output
• a boolean aigc_watermark flag that controls an AI-generated-content watermark

The image-to-video variant animates a still image into 2K video. It also accepts an optional last frame, which lets the model generate a transition from a chosen first frame to a chosen last frame. A reference-to-video mode is priced separately for output, video input and input images.

Spec	Value (Pika API)
Modes	Text to video, image to video, reference to video
Duration	4-15 seconds, default 5
Resolution	768P or 2K
Aspect ratios	21:9, 16:9, 4:3, 1:1, 3:4, 9:16
Seed range	-1 to 4294967295
Image-to-video extra	Optional last frame for first-to-last transitions

The sources differ slightly on clip length. Pika’s tagline describes “5–15s clips”, but its request schema accepts durations from 4 to 15 seconds with a default of 5. A separate third-party H3 site also cites a 4-15 second range for MiniMax’s official API. Builders should rely on the schema value.

How generation works: asynchronous jobs

Generation on Pika’s API is asynchronous. A POST to the generate endpoint does not return a video. It returns a job object with the status queued and an id. Clients store that id and poll the job until it reaches a terminal state, either completed or failed. A completed job carries its download URL in the output field. A separate content endpoint returns the same URL for clients that kept only the id.

Several details matter for production integrations:

• Deletion. A DELETE call erases the stored media and the request captured at submit, meaning the prompt and its parameters. After that, the content endpoint answers with HTTP 410.
• Usage and billing fields. Once a job finishes, it reports the units the provider measured, such as video_output_tokens, and a billing object with the charge in micro-dollars.
• Errors. Validation failures return HTTP 422 with a message that names the offending field. Rejections for balance or rate limits return a failed job envelope with an error code such as insufficient_balance.
• Idempotency. Reusing an Idempotency-Key header with a different request body returns a 409 conflict.
• Authentication. Requests authenticate with an X-API-Key header. Pika warns against exposing the key in browsers or other client-side code.

Pika says the API is plain JSON over HTTPS and needs no provider-specific SDK.

What Pika charges for H3

Price is part of this listing, so it is reported here as Pika states it. Pika lists text-to-video and image-to-video at $0.08 per second at 768P and $0.13 per second at 2K, and says only successful runs are charged.

Reference-to-video uses the same per-second rates for output and for video input. Input images cost $0.04 per image, with the first five included.

On the same page, Pika prices MiniMax’s older Hailuo 2.3 per request rather than per second. For example, Hailuo 2.3 costs $0.168 for a 6-second 768p clip and $0.294 for a 6-second 1080p clip. This change in billing model matters for budgeting. With H3, cost scales directly with clip length and resolution, so the 4-15 second duration setting affects spend as much as it affects output.

What the model is meant for, and what it is not

Pika’s documentation includes a short use guide. It lists H3 as best for:

• cinematic text-to-video shots
• short social and ad clips
• consistent multi-shot sequences
• look-dev and style exploration

The same guide lists H3 as not suited to:

• frame-exact rotoscoping
• feature-length renders
• pixel-perfect compositing
• precise on-screen text

These items come from Pika’s documentation.

That last limit is a practical warning for anyone generating ads or product videos. A third-party H3 site makes a similar point and advises that packaging text and factual product claims stay under human review, “because generative details can drift between frames”.

What other sources say about H3’s capabilities

Beyond Pika’s documentation, the remaining sources are third-party sites that sell access to H3. Their claims should be read with that in mind.

The site minimax3.com says MiniMax’s official V2 video API uses the model ID MiniMax-H3. It adds that third-party provider identifiers may differ. The site says the official API confirms multimodal generation from text, image, video and audio references.

The same site describes H3 as generating sound with the picture, including dialogue, ambience, music and timed effects. It also cautions that output-audio behaviour “can still vary by provider, prompt, and language”. Pika’s text-to-video request schema shown in the sources lists no audio parameter. Developers who need audio should confirm it with their provider.

Video platform Vidnix announced H3 support in a post dated July 31, 2026. Vidnix describes H3 as a multimodal model that understands text, images, audio and video together. It lists image-to-video, video-to-video transformation, motion transfer and creative visual editing as supported workflows. That post also claims smoother motion and faster generation, but gives no measurements to support either claim.

What it means for builders

For teams already building on video APIs, Pika’s H3 listing offers several useful features:

• a stable async job pattern
• six aspect ratios covering widescreen through vertical social formats
• a seed parameter for repeatable runs
• explicit deletion of stored prompts and media

These details come from Pika’s API reference. The first-to-last-frame option in image-to-video is useful for storyboard-driven work, where both ends of a shot are already approved.

Open questions remain:

• No independent quality evidence. None of the sources include third-party benchmarks or leaderboard results for H3. Quality claims come from resellers’ marketing copy.
• Audio is unclear. Native audio is described by a third-party site but not exposed in Pika’s documented request parameters.
• Specs vary by provider. The minimax3.com site advises validating authentication, rate limits, safety behaviour, file constraints, pricing and actual output settings before treating any integration as production-ready.
• No open weights. None of the sources describe downloadable weights or a licence for H3. It appears only as a hosted model.

What to watch next

The next useful signals are:

• independent evaluations of H3 against other video models, such as human-preference leaderboards
• whether Pika exposes audio controls in its H3 endpoints
• whether MiniMax or its distribution partners publish technical details that would let builders compare H3 with other models on more than marketing copy

Until then, the strongest guide to H3 is Pika’s own documentation: what the parameters allow, what the model is not suited for, and what each second of 2K output costs.