Pika H3 Flex: Video API Pricing and Local Setup

Pika’s November 7, 2026 price table bills text-to-video and image-to-video by output second. The amounts below use exactly 15 seconds; native frame rounding can make the final clip slightly longer. Reference-to-video has a higher output rate and separately bills input-video seconds. Use the mode you actually submit when calculating a project.

Create a Pika API account and key, arrange billing, and set PIKA_API_KEY in your local or server environment. This request asks for 15 seconds at native 768p. It creates a queued job; save its returned id and poll until completed, failed or cancelled.

The downloadable release produces 5–15-second video at native 480p or 768p, 24 fps, with 32 kHz stereo audio. Its six-step schedule belongs to the released adapters. FL2VA handles text and endpoint frames; Ref2VA uses the separate reference transformer and adapter. Match the workflow and adapter when loading weights.

Pika’s API accepts 3–15-second requests. Its 1080p, 2K and 4K options generate at 768p and then upscale. Hosted finishing, prompt enhancement and audio separation are separate capabilities from the local adapters. A requested duration may round up to the supported frame grid.

The launch’s seven-second figure measures DiT computation for a 15-second 768p example. Loading, encoding references, decoding and service overhead affect complete clip time. Use end-to-end job time and the cost of retained clips for a production comparison.

Pika’s adapters inherit the MiniMax H3 Community License. Its standard grant excludes the United States, European Union, United Kingdom and South Korea; deployment in an excluded territory requires a separate license. The agreement also covers use through hosted services. For API access, establish the terms applicable to your account and location before choosing that route.

When your commercial products and services generate more than $20 million in yearly revenue, the agreement requires separate prior written authorization. Apache-2.0 licensing for example code or the text encoder does not replace the model-weight agreement.

Pika’s tested CLI uses one GPU per request, with representative runs on H100 80GB. Default 480p five-second sampling allocated about 66–69 GiB. At 768p for 15 seconds, the default exceeded H100 memory; CPU offload brought sampling allocation to about 29–32 GiB GPU memory while observed host memory approached 94 GiB, including page cache.

Sampling figures exclude the text encoder and VAE stages. Allow additional headroom for references, buffers and dependencies. The guide asks for at least 211 GB for base components plus about 5.5 GB for both adapters, Python 3.12, a supported NVIDIA driver and pinned dependencies. These are tested configurations, not minimum-memory promises.

Install git, patch, ffmpeg and the C compiler development tools required by the official README. Then choose the repository’s pip, Conda or Docker installation. The CLI downloads the pinned base components and selected adapter. For ComfyUI, use the Pika-specific workflow and matching LoRA. For Diffusers, follow the documented ModularPipeline integration rather than a generic image-generation snippet.

Keep the official six-step package separate from Turbo LoRA experiments. An adapter trained for the reference transformer cannot be substituted into the endpoint-frame transformer. Record the adapter, base revision, precision, duration and reference set with each result.

Reddit author Important-Apple7422 reports using an RTX 4090 24GB with 64GB RAM for four reference-guided, 15-second Pika H3 Flex clips. Each reportedly took about 40 minutes. The post links its workflow separately but does not state the exact weight precision or peak memory. Treat it as a configuration-specific runtime example, rather than a comparison with Pika’s hosted service or official CLI tests.

Choose Pika when its frame or reference controls and finishing route fit your shot. Compare MiniMax H3 when you need the base pipeline or another serving provider. Gemini Omni 1.2 Flash offers a different workflow built around follow-up conversational edits. Match input mode, duration, audio and output resolution before comparing prices.

For a hypothetical rented GPU at $3 per hour, a complete ten-minute attempt costs $0.50 in allocated compute. Three attempts per retained clip cost $1.50 before loading, idle time, storage and finishing. Replace those assumptions with measured wall time; a six-step count alone cannot establish that local generation is cheaper.