What Pika actually shipped
The useful part of this release is the workflow. Video generation is an iteration problem: the first clip is a draft, and the real cost is the versions you throw away. Pika post-trained MiniMax H3 with that in mind. It fine-tuned the model on its own curated audio-video data, distilled the 50-step sampling schedule down to six steps, then trained reward and preference adapters and merged them into the six-step model at no extra inference cost.
The result takes text, an opening or closing frame, or a stack of references, and returns clips of five to 15 seconds at 24 frames per second with synchronized stereo audio. The reference mode is the professional feature: up to nine images, three videos and three audio clips — 12 reference files in total — so a character sheet, a location and a movement reference can go into one generation. That is a real change in how a shot gets directed. It is also where the limits sit, and Pika publishes most of them. The rest of this piece is those limits, because they decide who this release actually serves.
Seven seconds, with the clock defined
"Faster than real time" is the headline, and it comes with a footnote Pika wrote itself. The seven seconds for a 15-second 768p clip measures transformer computation in a controlled benchmark. It is not end-to-end API latency. The technical deep dive is more precise still: its transformer comparisons against fal's hosted H3 models exclude conditioning, decoding and serving overhead, and even its warm-inference numbers, which do include conditioning and decoding, exclude serving overhead.
None of that makes six steps slow. Cutting 50 steps to six is a structural saving, and it will show up in any honest end-to-end measurement too. But the number a creator feels is the time from pressing generate to watching a clip, including queueing, decoding and the finishing pass. Pika has not published that number. Read the seven seconds as transformer computation time, not as a promise about your API call and not as a local benchmark. Separately, the technical deep dive's hosted inference-acceleration measurements were conducted on eight B300 GPUs. Locally, the model card's tested configuration is an H100 with 80GB of memory, running one request at a time.
Open, in two parts
The open-weights release is genuine, and narrower than the phrase suggests. What you download from Pika is two adapters: FP32 LoRA packages for the frame-conditioned and reference-conditioned tasks, plus inference code and a pinned sampling schedule. The base they sit on is MiniMax H3, and the package bundles its components in Diffusers format at about 210 GB in total. A single workflow loads only its own partition of that, but this is not a laptop download, and the card is plain that maximum-size reference combinations are input limits, not a guarantee that they fit in GPU memory.
The license is the second narrowing. The adapters are model derivatives of H3, distributed under the MiniMax H3 Community License Agreement. That agreement covers the world except the European Union, the United Kingdom, South Korea and the United States, and it bars displaying the works or their outputs outside that territory. A creator in Canada falls within the agreement's Applicable Territory. A creator in the United States does not have permission to use the downloadable weights under this community license. Pika's hosted API is a separate access route, but its availability should not be treated as a blanket exception to the underlying license restrictions. Commercial services exceeding US$20 million in yearly revenue also require prior written authorization under the agreement. This is the base model's license traveling with the derivative, which is why the license line belongs in any coverage of an H3 derivative.
The third narrowing is the product split. Three launch features are not in the local package. The 4K finish is an upsampler: Flex generates natively at 480p or 768p, and Pika's separate 1B-parameter upsampler, served through its API, takes the chosen draft to 1080p, 2K or 4K. The technical post says Pika will open-source that upsampler, in the future tense and with no date. The five separate audio stems and prompt enhancement are likewise API endpoints; locally you get one mixed stereo track. So "open model, 4K, editable audio" describes the combined hosted-plus-local product. Each part is available. They are not all available in the same place.
Starting price means starting
Pika's pricing tables are unusually concrete, and they reward reading both. Text- and image-to-video starts at $0.01 per output second, against $0.08 for MiniMax H3 and $0.0325 for fal H3 Max in the same table. Reference-to-video, the mode the launch is really selling, starts at $0.03 per second, three times the headline rate. A 15-second reference draft at the starting rate is $0.45, before the finishing pass. That is still cheap for iteration, which is the point of the model. It is just not the cent-a-second figure the first table leads with, unless your work needs no references.
The release pairs "1.8x faster than fal H3 Max" with "up to 80% cheaper," but does not identify the pricing comparison behind the latter figure. Using Pika's published starting rates, Flex is about 69% cheaper for text-to-video and 40% cheaper for reference-to-video. Both are meaningful reductions; neither reproduces the 80% headline. Until Pika specifies the comparison, the published tables are the figures buyers can independently check.
What the scores do not settle
Pika's evaluation is its own: same prompt set, 768p, four systems. Flex took the highest score on all five visual and audio quality metrics, while alignment between audio, video and text stayed comparable across the four. Its human preference study put Flex first on anatomy, composition and audio sync, and other systems first on physics, multi-scene narrative, lighting and materials, and action-heavy reference tasks. The DesignArena ranking, first among open-weights models and seventh overall, is a launch snapshot, in Pika's own words.
Pika's conclusion then says the quiet part itself: physics, multi-scene consistency and complex reference control remain areas for further improvement. Those are not edge cases for professional work; they are most of the notes a director gives. Add the model card's warning that its showcase clips are not release-validation renders and that exact reproduction with the released package is not guaranteed, and the honest summary is narrower than the launch: a much faster H3 with strong measured draft quality, whose weakest areas are the ones iteration is meant to fix. Six-step drafts at these prices make that trade workable. You can afford to discard the versions where the physics fails. That, rather than any single benchmark, is the case for Flex.
Sources
- [1] Pika, "Introducing Pika H3 Flex" (launch post, October 29, 2026) — six steps vs. 50, 12 references, five audio stems via the API, native 480p/768p with hosted upscaling, adapters and inference code release, API Club now and creation product on November 2, and the footnote defining the seven-second figure as DiT computation in a controlled benchmark, not end-to-end latencyRead source
- [2] Pika Research, "Pika H3 Flex Technical Deep Dive" (October 29, 2026) — three-stage post-training (SFT, PDMD, ReFL/DPO LoRA), pricing Tables 1 and 2 ($0.01 text/image, $0.03 reference; comparators), latency methodology (transformer comparisons exclude conditioning, decoding and serving overhead; acceleration measured on eight B300 GPUs), its own quality and human-preference results including the categories other systems led, the 1B upsampler and "We will opensource the upsampler model," and the conclusion naming physics, multi-scene consistency and complex reference control as areas for further improvementRead source
- [3] Pika-Labs, Pika-H3-Flex model card, Hugging Face — input/output specifications, reference limits (9 images / 3 videos / 3 audio / 12 total), adapter packages (FP32 LoRA rank/alpha 128), base components at approximately 210 GB, one request on one GPU, representative tests on H100 80GB, hosted API extensions listed separately, showcase reproduction caveat, and the license statement (Model Derivatives under the MiniMax H3 Community License Agreement)Read source
- [4] MiniMax H3 Community License Agreement, MiniMax H3 repository, Hugging Face — Applicable Territory definition excluding the EU, UK, South Korea and the US (Definitions 3 and 5), section V.4 territory restriction on use and display of outputs, section IV.1 separate authorization above US$20 million yearly revenueRead source
- [5] Business Wire, "Pika Launches Pika H3 Flex, a Video Model Built for Professional Iteration" (October 29, 2026, 18:40 UTC) — DesignArena #7 overall claim, "1.8x faster than fal H3 Max, and up to 80% cheaper" line, launch offer (first two weeks: 15 free generations per day, all generations 50% off), access routes and November 2 creation-product dateRead source