Pika H3 Flex review: price, licence limits and hardware

This review is built from what could be checked on November 3, 2026: Pika's launch post, technical deep dive, H3 Flex page, API documentation, pricing and terms; the Hugging Face model card and GitHub repository; the MiniMax H3 Community License; DesignArena's text-to-video leaderboard; ComfyUI's tutorial; local runs published by two Japanese blogs and by the makers of a Mac port; and reporting by AlphaSignal and i-SCOOP. Speed, quality and ranking figures are Pika's own claims unless another source is named.

What Pika H3 Flex is

Pika H3 Flex is a video-and-audio model from Pika, the AI video company run by Mellis, Inc. of Palo Alto, California (terms of service), founded in 2023 and backed by Lightspeed (press release). Flex is not a new base model. Pika took MiniMax H3, the 33-billion-parameter transformer that MiniMax published with open weights in the summer of 2026 (MiniMax H3 model card), fine-tuned it on clips it curated, distilled it from 50 sampling steps to six and tuned it further with reward learning (technical deep dive). What Pika ships is two LoRA adapters that sit on top of MiniMax's base files: FL2VA for text and first- or last-frame generation, and Ref2VA for clips built from references (model card).

Next to its base model, from the two model cards and the two makers' price lists:

Sources: Pika H3 Flex model card, Pika H3 Flex page, MiniMax H3 model card, MiniMax pricing, DesignArena.

What changed on October 29

Pika released H3 Flex on October 29, 2026 through its developer API, the Pika API Club, with the weights on Hugging Face, the code on GitHub and ComfyUI templates the same day (launch post, Pika on X). For the first two weeks Pika halved its API prices, and API Club members got 15 free generations a day from October 29 to November 5 (press release, H3 Flex page).

The launch post said Flex would reach the Pika creation app on November 2. On November 3 the model menu on pika.art and the app's pricing page did not list it, and the H3 Flex page still named the API Club as the place to try it.

Two launch claims have already moved. The launch post and the deep dive promised five audio stems (dialogue, score, SFX, foley and ambience); the model page and the API documentation now return three, Dialogue, Music and Effect, and the page says its interactive demo uses "the earlier five-stem preset". Pika launched at 7th place and 1293 Elo on DesignArena's text-to-video board; on November 3 the board showed 8th place and 1288.

How fast it is, and what "7 seconds" means

Pika's headline is a 15-second 768p clip with sound in as little as seven seconds. Its own footnote says this is transformer (DiT) computation in a controlled benchmark, not the time an API request takes: references, upscaling, stem separation and service overhead add to it. The acceleration figures in the deep dive were measured on eight Nvidia B300 GPUs, where a 5-second 768p clip took 2.129 seconds of full inference after Pika's optimisations, down from 3.409. Pika also claims Flex is 1.8 times faster than fal's H3 Max, averaged across text, image and reference tasks (deep dive).

DesignArena publishes its own timing column. On November 3, 2026:

Source: DesignArena video leaderboard, updated November 3, 2026 at 04:08 UTC. Flex ran about 19 times faster than MiniMax's own H3 there, which is the point of the release. The arena does not show the lead over H3 Max that Pika claims: fal's H3 Max averaged 4.6 seconds. DesignArena does not say which resolution or length each model ran at, so treat the column as a rough comparison. On your own hardware the numbers are minutes, not seconds.

What it costs

Pika sells H3 Flex by the second of output video. Prices on the H3 Flex page on November 3, 2026, in USD per second:

With a video reference, the input seconds are billed at the same rate as the output; image and audio references are free. Finishing a 480p or 768p draft at a higher resolution charges only the difference, audio stems add 0.001 USD a second, and the prompt rewriter is free on Pika's hosting but 0.02 USD a request for anyone running the weights. Only successful generations are billed (API documentation).

At regular prices a 15-second text-to-video clip costs 0.60 USD at 768p, 1.05 USD at 1080p and 2.40 USD at 4K. A 15-second 768p clip guided by a 10-second video reference costs 3.00 USD, because 25 seconds are billed at 0.12 USD. MiniMax charges 0.08 USD a second for H3 at 768p on its own platform, so Flex through Pika costs half as much at regular prices (MiniMax pricing).

The API needs an API Club membership of 10 USD a month, which comes back as 10 USD of API credit each month, and a prepaid balance; Pika's documentation says there is no free tier for generation (API pricing, llms.txt).

The Pika app has its own plans, separate from the API:

Yearly billing takes 20% off. Extra credits cost 60 per US dollar, expire after 365 days and cannot be spent on the API (Pika pricing). Pika had not published a credit price for H3 Flex on November 3, 2026; for scale, its pricing page puts a 5-second 720p Seedance 2.5 clip at 122 credits and a Wan 3.0 clip at 33. Sites named after the model sell their own H3 Flex credits, and both pikah3flex.org and pika-h3.com say in their footers that they are not affiliated with its makers.

Running it on your own hardware

The open release covers text to video, first and last frame, and reference to video at 480p and 768p with mixed stereo sound. The upsampler to 1080p and above, the stem separator, the prompt rewriter and Pika's fast serving stack stay on Pika's servers; Pika says it will open-source the upsampler later (launch post, deep dive).

Pika's own package is demanding, according to its hardware notes:

• about 211 GB of MiniMax H3 base files plus about 5.5 GB of adapters on disk;
• one Nvidia GPU per request, with no way to split a clip across several cards;
• 66 to 69 GiB of GPU memory for a 480p, 5-second clip on an H100 80 GB;
• CPU offload for a 768p, 15-second clip, which then used 29 to 32 GiB on the GPU and about 94 GiB of system memory;
• Python 3.12, with H200 and B300 cards not yet qualified.

Community builds bring it to consumer cards. ComfyUI's templates pair compressed int8 base files with the Flex adapters, and ComfyUI says a 5-second 1344×768 clip finishes in about a minute on a single card, without naming the card (ComfyUI tutorial). The public runs are slower:

Sources: TrueNorthAI, Rossy Laboratory, Phosphene on GitHub. Each is one setup, published between October 30 and November 1, 2026. Before downloading anything, read the licence section below: in four large markets the licence does not allow running the weights at all.

What the public tests and users report

Pika's own evaluation. At 768p on one prompt set, Pika scores Flex highest on all five visual and audio quality metrics against MiniMax H3, fal's H3 Max and HeyGen Video, but the margins are small: video aesthetics 5.41 against 5.28 for MiniMax H3, and the same text-video alignment score as MiniMax H3, 0.225 (deep dive). In Pika's blind human comparisons Flex led on anatomy, complex composition and audio sync, and on style, framing and camera motion for reference clips; the other models were preferred on physics, multi-scene narrative, lighting and materials, and action-heavy or video-controlled references. Pika's conclusion names physics, multi-scene consistency and complex reference control as areas still to improve.

DesignArena. Voters compare two clips made from the same prompt without knowing which model made which, and the board ranks models by Elo. Flex's 1288 from 817 votes puts it level with Kling 4.0 and 12 points above MiniMax H3, and its 59.5% win rate carries a margin of plus or minus 3.4 points (DesignArena). It remains the highest-rated open-weight model on the board. explainx.ai, which ran no tests of its own, warns that a gap of a few Elo points means little.

ComfyUI. Comfy's tutorial, published with the release, says camera moves are advisory: explicit degrees or cut timestamps can land 0.2 to 0.3 seconds off or go the wrong way, a requested wide shot tends to come out as a medium shot, and a hand told to stop just short of something often touches it. It also warns against stacking a turbo LoRA on the adapter, which breaks the output (ComfyUI).

Users. TrueNorthAI found Japanese pronunciation and movement natural in a 15-second spoken clip, rated realism 4 out of 5 with the compressed base files, and saw weaker skin texture at 15 seconds than at 5. In a reference test it had to state that every view on a character sheet was the same person (TrueNorthAI). Rossy Laboratory praised how Flex kept several timed events in order in a 5-second station scene, but a giant goldfish shot looked like CG even after the prompt was rewritten for live action, and the blog stayed with its existing accelerated H3 setup (Rossy Laboratory). The makers of the Mac port found that people in the first frame who are not named in the prompt tend to vanish, and that their 15-second clips, stitched from three 5-second windows, let sets and side characters drift (GitHub issue). On Hugging Face one user reported bad results when pairing the adapter with pruned bf16 base files (discussion). Most launch coverage, including AlphaSignal's, repeats Pika's own benchmarks.

Licence, privacy and content rules

The adapters are "Model Derivatives" of MiniMax H3 and come under the MiniMax H3 Community License; only the example code is Apache 2.0 (model card). The licence, dated August 2, 2026 and issued by Nanonoble Pte. Ltd., the MiniMax company named in it, says (licence text):

• it applies worldwide except the European Union, the United Kingdom, South Korea and the United States, and neither the weights nor their outputs may be used there;
• a commercial product earning more than 20 million USD a year needs MiniMax's written approval;
• a commercial product must show "MiniMax H3" prominently in its interface;
• outputs may not be used to improve any other AI model;
• AI-generated content posted publicly must be clearly marked as machine-generated, and military use is banned;
• disputes go to the courts of Hong Kong.

Pika's launch pages do not mention the excluded territories. i-SCOOP pointed out on November 3 that Europe can therefore only rent the model through a hosted API, and that it is not public whether Pika's arrangement with MiniMax covers European customers; it advises asking Pika in writing (i-SCOOP). Pika is based in the US and sells the hosted model there, and neither company has published the terms that allow it.

On Pika's own service you keep your inputs and outputs, and commercial use depends on your plan: the free app plan has no commercial licence (terms, pricing). Pika's privacy policy, last changed on October 21, 2026, lets it use inputs and outputs to improve its models where the law allows or you consent, and says personal information sent through the API "will generally be used to improve and train" AI models; an enterprise agreement can set other terms (privacy policy). Subscription payments are not refunded.

Who should not use Pika H3 Flex

• Anyone in the US, EU, UK or South Korea who wants to download and run the weights or ship them in a product: the licence does not cover those places or the outputs made there.
• Anyone without an 80 GB-class Nvidia GPU who needs Pika's official package at 768p and 15 seconds; compressed community builds run on a 4090 or 5090 at minutes per clip.
• Anyone who needs native 1080p detail, clips longer than 15 seconds, or reliable physics and action: Kling 4.0 makes up to 30 seconds, and Pika's own evaluation puts other models ahead on physics.
• API customers who cannot let their inputs train models and have no enterprise agreement with Pika.
• Companies earning more than 20 million USD a year that plan to build on the weights without MiniMax's approval.
• Pika app subscribers who need a known credit cost today: Pika had not published one on November 3, 2026.

Alternatives

Prices from each maker's own price list on November 3, 2026, and Elo from DesignArena's text-to-video board on the same day:

Sources: MiniMax, Kling AI, BytePlus, Alibaba Cloud, Google, DesignArena.

MiniMax H3 is the better choice if you want MiniMax's own 2K regeneration and hosted prompt processing, or to buy from the model's maker; MiniMax's lighter H3-Lite costs 0.04 USD a second at 768p, the same as Flex's regular price. It comes under the same licence, so it does not help anyone in the excluded territories.

Kling 4.0 is the better choice for long takes: up to 30 seconds in one generation at native 1080p (Kling 4.0 guide). It ties Flex on DesignArena but took about 25 times longer per clip there, and costs about four times as much a second at 720p as Flex does at 768p.

Seedance 2.5 from ByteDance sells through BytePlus with native 1080p and a 480p draft mode. It sits 34 Elo points below Flex on DesignArena and costs almost six times more a second at 720p than Flex at 768p, so pick it only if its 1080p output wins on your own material.

Wan 3.0 from Alibaba sits 33 points above Flex on DesignArena and makes native 1080p at a list price of 0.20 USD a second; Alibaba Cloud showed a limited-time 30% discount and a 30-second free quota in its Singapore region. It is the better choice for quality at 1080p when you do not need open weights.

Gemini Omni 1.1 Flash tops DesignArena at 1360 and costs about 0.10 USD a second at 720p through the Gemini API; Google's older Veo 3.1 sits far lower on the same board, at 1164. Omni Flash is the better choice when quality matters more than price or openness.

If you mainly want speed from H3 through a host, fal runs its own H3 Max variant, which DesignArena timed at 4.6 seconds a clip and rated 1277. If the territory clause rules Flex out, check the licences of the other open models on DesignArena's board, such as Lightricks' LTX 2.5 versions at 1163 and 1157 Elo.