Pika H3 Flex Is Out: Features, Pricing and Access

What is Pika H3 Flex?

Pika H3 Flex is a video generation model from Pika Labs, built by post-training MiniMax H3. Pika announced it on October 29, 2026 as a model "built for professional iteration," with open weights, up to 12 references, separate audio stems and output up to 4K.

Flex is not trained from scratch. According to Pika's announcement, the team post-trained MiniMax H3 to generate in six steps instead of 50, then refined the faster model for quality and creative control. The Hugging Face card tags it as an adapter on MiniMaxAI/MiniMax-H3.

Pika frames the model around fast iteration: in its words, the first generation is a starting point, and the creative work is what happens next.

Pika H3 Flex release date and availability

Pika H3 Flex was announced on October 29, 2026 and is live on the Pika API. According to Screenweaver, it reaches the pika.art creation app on November 2. The weights are on Hugging Face and GitHub.

Screenweaver says the announcement also went out as a Business Wire press release. Pika's own account posted on November 4 that the model debuted as the #1 open-weight model on DesignArena's text-to-video board.

Outside developers have already built on it. In the same post, Pika says @_akhaliq built a Turbo LoRA less than 24 hours after launch, with 2-step draft generation plus 4-step fast and 8-step HD variants, compatible with ComfyUI and Diffusers. That is Pika's description of a third-party release, and we have not seen the release itself.

Pika H3 Flex key features and specs

Flex supports text to video, first and last frame, reference to video and reference-guided editing. Clips run 5 to 15 seconds at 24 fps, with native 480p or 768p output and hosted upscaling to 1080p, 2K or 4K. Specs below come from Pika's page as read by Screenweaver on November 1.

Reference items for a product film laid out on a studio table

Some rules matter in practice. At least one image or video reference is required, so audio-only requests are refused. Video and audio references each run 2 to 15 seconds. The per-type caps add up to more than 12, so you trade slots between types.

The open release is narrower than the hosted one. Per the Hugging Face card, local inference covers native 480p and 768p with mixed stereo audio, while 1080p to 4K, audio stems and prompt enhancement are Pika API extensions.

Pika H3 Flex speed claims and benchmark ranking

Pika says Flex can generate a 15-second 768p clip with synchronized audio in as little as 7 seconds of transformer computation. That is a controlled benchmark of transformer time, not the wait you will see on an API call, and Pika's page says so.

Pika also claims Flex is up to 1.8 times faster and up to 80% cheaper than H3 Max, averaged across text, image and reference jobs.

On quality, Pika says Flex ranks #1 among open-weights models and #7 overall on DesignArena's text-to-video leaderboard. Screenweaver reports a score of 1293, above base H3, H3 Max and Seedance 2.5 according to Pika. That is one crowd-voted board. Run your own product shots before moving a campaign over.

Pika H3 Flex pricing and free access

Pika has confirmed relative savings (up to 80% cheaper than H3 Max) and, per Screenweaver, $0.001 per second for audio stems. A full per-second video price was not available to us. A Reddit post claims about $0.01 per second, roughly $0.10 for 10 seconds, but that is not confirmed by Pika.

The weights are published on Hugging Face and GitHub. In the app, Pika's pricing page lists these plans:

Credit cost per Flex generation is not published in the pages we reviewed. Pika's pricing page shows all models are included on every plan.

How to use Pika H3 Flex for AI video and static ads

Flex fits ad teams that iterate on one product across many angles. Pika's own example turns a single product reference into a multi-shot lip-gloss commercial, with the reference guiding the product's look while setting, camera and style change.

Product video variations lined up on an editing timeline above separate audio tracks

Where it could slot into a workflow:

• Hook variants: a 5 to 15 second clip with fast generation lets you test several openings per concept. See our hook rate and hold rate benchmarks for what to measure.
• Product b-roll: feed up to 9 product photos and a motion video reference, then change setting and camera.
• Localization and edits: with dialogue on its own stem, you can mute a generated line and lay a recorded or translated one under the same picture.
• Format coverage: 11 aspect ratios from 21:9 to 9:21 cover vertical and widescreen placements.

Screenweaver notes Pika does not publish a test of how well a voice sample holds across separate generations, so check that before building a spokesperson series on it. For comparing Flex with other models side by side, the AdGraam Playground is one place to do it, and Flow Builder can chain generation and upscale steps. For broader context on AI UGC, see our guide to AI UGC ads.

Pika H3 Flex vs other AI video models

Flex is closest to its parent, MiniMax H3, and Pika says it beats H3, H3 Max and Seedance 2.5 on one DesignArena board. The clearest practical difference is the reference budget: 9 images against the 30 Screenweaver says Seedance 2.5 accepts.

Related open work is also appearing on the same base: see P-H3-Image from Pruna.

What to watch next

Watch for three things: a published per-second API price, credit costs inside the pika.art app, and the open-weight license terms for commercial use. Until those are out, you cannot work out what Flex costs at ad volume.

Also watch the community adapters. Pika highlights a Turbo LoRA that could make drafts faster still, but its quality is unverified. Side-by-side tests on product shots, and on whether a voice stays the same across separate clips, are still missing.