Pika H3 Flex: What It Is, What It Costs, and How It Compares to MiniMax H3 Max

Short answer: Pika H3 Flex is the model Pika announced on October 29, 2026: a post-trained MiniMax H3 that makes a 5 to 15 second clip with stereo sound in six sampling steps, takes up to 12 reference files, and is published as open adapters on Hugging Face. Pika's API bills it from $0.01 per output second at 480p. Cicadas does not offer Flex. It runs the MiniMax H3 Max version that fal hosts, from 63 credits for 5 seconds, which is about $0.62 on the first credit pack.

We make Cicadas, an independent site; we have no tie to Pika or MiniMax. We have not run Flex anywhere, not in Pika's app, not through its API and not on a GPU of our own. Everything below about Flex is what Pika printed on its own blog, API reference, price page and Hugging Face model card on November 3, 2026, and each source is listed in our research note, docs/research/claude-pika-h3-flex-20261103/sources.md. Where Pika's pages disagree with each other, we say so.

1. What Pika H3 Flex is

• A post-trained MiniMax H3. Pika's announcement calls Flex "our post-trained MiniMax H3 model". Pika cut sampling from the base model's 50 steps to six, then tuned the faster model for quality. On Hugging Face the model is listed as an adapter of MiniMaxAI/MiniMax-H3, not a new architecture.
• Two adapters. FL2VA handles text alone, a first frame, a last frame or both; Ref2VA handles mixed references. Both are LoRA files of rank 128 that load on top of the MiniMax H3 base.
• References. Ref2VA takes up to 12 files at once: at most 9 images, 3 videos and 3 audio clips, with at least one image or video. Each video or audio clip must run 2 to 15 seconds.
• Output. 480p or 768p natively, 5 to 15 seconds (Pika's hosted text-to-video page says 3 to 15 seconds), 24 frames per second, in 11 aspect ratios from 21:9 to 9:21, with 32 kHz stereo sound mixed into an MP4. Prompts can run to 8,000 characters. A 5-second request actually returns 124 frames, about 5.17 seconds, because length snaps to the model's frame grid.
• Speed claim. Pika says a 15-second 768p clip can take "as little as seven seconds" of transformer computation, and its own footnote says that is a controlled benchmark, not the time an API request takes from start to finish.
• Ranking claim. At launch Pika said Flex was first among open-weight models and seventh overall on DesignArena's text-to-video leaderboard. That is Pika's snapshot from October 29; leaderboards move.

2. What "open weights" covers, and what it does not

The public release is the two adapters, the inference code and a bundle of MiniMax H3 base components. The adapters are "Model Derivatives of MiniMax H3", so they carry the MiniMax H3 Community License Agreement; the Apache 2.0 license on the repository covers the example Python code only, not the weights or the sample media. Read the MiniMax agreement before building a product on the weights.

Running it yourself is a data-centre job. The model card puts the base components at about 210 GB, and Pika's hardware notes on GitHub say representative 480p and 768p requests passed on an H100 with 80 GB of memory, one request per GPU. Three things never leave Pika's servers: the upscaler that finishes a clip at 1080p, 2K or 4K, the audio-stem separation, and the optional prompt rewriter. Run locally, Flex stops at 768p with one mixed soundtrack.

3. Where you can run it

• Pika's API. Text to video, image to video, reference to video, a draft-to-4K finishing step, stems and the prompt rewriter, behind one API key. The announcement points first to the Pika API Club.
• Pika's creation app. The announcement says Flex was "coming to the Pika creation product on November 2". We have not signed in to check that it arrived.
• Your own hardware. Weights on Hugging Face (Pika-Labs/Pika-H3-Flex), code on GitHub, three ComfyUI workflow templates (reference, first and last frame, text) and a Hugging Face Diffusers pipeline.
• Not on fal.ai, not on Cicadas. fal's model search returned no Flex endpoint on November 3, 2026, and every Flex row on Pika's own price page says "Only on Pika".

4. What Pika charges

These are Pika's API rates per second of output, read on its price page on November 3, 2026:

• Text or image to video: $0.01 at 480p, $0.02 at 768p, $0.035 at 1080p, $0.04 at 2K, $0.08 at 4K.
• Reference to video: $0.03 at 480p, $0.06 at 768p, $0.075 at 1080p, $0.08 at 2K, $0.10 at 4K, and reference videos are billed per second at the same rates.
• Extras: audio stems $0.001 per second, prompt rewriter $0.02 per request.

Worked through: 5 seconds from text at 480p is $0.05; 15 seconds at 768p is $0.30; the same 15 seconds finished at 4K is $1.20. These are developer prices for an API you call from code. We did not read the plans inside Pika's consumer app, and some third-party articles quote other figures that Pika's page does not show, so we left those out.

One mismatch on Pika's side: the announcement lists five stems (Dialogue, Score, SFX, Foley and Ambience), while the API reference we read describes the stem option as returning Dialogue, Music and Effect files. If stems matter to you, test the endpoint before you plan around five tracks.

5. Flex next to the MiniMax H3 Max that Cicadas runs

Cicadas reaches MiniMax H3 Max through fal's minimax/h3-max endpoints on the MiniMax H3 page, which Hailuo AI also opens. That is the H3 Max version fal hosts; Flex is Pika's post-training of MiniMax H3. Same family, different tuning, and we have only filmed one of them.

What the H3 Max form on that page does: text to video or one photo to video; 5, 10 or 15 seconds; 480P or 768P; 16:9 or 9:16 for text prompts; sound on every clip, with no silent mode. It does not give you stems, 4K or a reference mode that mixes images, video and audio. Quotes, at the $9.90 first pack of 1,000 credits:

• 480P: 5 seconds 63 credits (about $0.62), 10 seconds 125 (about $1.24), 15 seconds 188 (about $1.86).
• 768P: 5 seconds 100 credits (about $0.99), 10 seconds 200 (about $1.98), 15 seconds 300 (about $2.97).

Per second, Pika's API is cheaper, and the gap is real. The comparison is not like for like, though: Pika's rates are wholesale prices for developers who write code against an API, while a Cicadas credit buys a finished clip from a web form, kept privately in My videos, with credits returned when a job fails. Pika's own comparison table also lists fal's H3 Max from $0.0325 per output second next to Flex at $0.01; that is Pika's figure for fal, which we have not checked against fal.

Here is a recorded H3 Max clip, so you can see the family's 480P output. It is not Flex. Text to video, 5 seconds, 480P, vertical, generated on September 29, 2026 (fal request 01a0ef93-6ebd-70d1-b346-778ed6261de0):

The prompt asked for a high view of a rainy crossing at night with umbrellas and headlights, a static camera and no text. The crowd and the wet road hold up; the shop signs came back with lettering nobody can read. Some of Flex's showcase prompts ask for words such as MOVE and SUNDAY, but Pika notes that its showcases are not release-validation renders and may not reproduce, so we would not assume Flex fixes lettering until we see it.

6. Which one to use

• Choose Flex if you call video models from code, want up to 12 references in one shot, need separate dialogue and music tracks, or want a 4K finish. Choose it too if you have the GPUs to run open weights and accept the MiniMax license.
• Choose H3 Max on Cicadas if you would rather type a prompt into a form, see the exact credit quote before you generate, and keep the MP4 without setting up an API key. The MiniMax H3 prompt guide has templates taken from our own recorded clips, and the MiniMax H3 Max guide covers the settings.

Credit packs and plans are on the pricing page. If Pika's numbers or availability change, the research note is where we will record it first.