Genativity

news

Pika's Enact preview publishes three real-time speed figures measured three different ways

The research preview reports 1.8x, 1.75x and 1.50x real-time generation across two different hardware setups, discloses only one resolution, and is not available to anyone outside enterprise conversations.

By Daniel Ochoa ·

Pika announced Enact on November 18, 2026, a research preview built on Pika H3 Flex that lets a creator drag actors, speak notes and move the camera while a scene plays. Pika reports 1.8x, 1.75x and 1.50x real-time generation, measured on different hardware, and the preview is not publicly available.

Pika's Enact announcement carries more performance detail than most AI video releases, and the three headline numbers in it are not measured the same way. Live Staging is reported as a response latency on a single H100. Live Camera Control is reported as time to first frames on a node with eight NVIDIA B300 accelerators. Live Actor Direction carries a throughput multiple and no latency figure at all. All three come from Pika, on Pika's hardware, and none of them can be reproduced by anyone else, because Enact is a research preview that is not available for consumer use.

The preview went up on November 18, 2026 as a post on Pika's research blog and a video from the company's X account. Enact is described as an experimental live-streaming environment rather than a model, built on Pika H3 Flex, the open-weight video model the company released in late October.

Three interaction modes with three incompatible measurements

Live Staging lets a creator grab a person or object with the mouse and move it while the scene continues, with the model adjusting pose, lighting and shadow around the new position. Pika puts it at 1.8x real time with under 0.4 seconds of response latency on one H100, including decoding in low-latency mode. Live Actor Direction runs a script model that turns a spoken note into structured beats and a real-time video model that performs them, at 1.75x real time for picture and sound together. Live Camera Control renders 640x480 at 24 frames per second on the eight-B300 node, reaching first frames in 0.51 seconds and then producing 0.473 seconds of compute per 0.708-second chunk, which is where the 1.50x figure comes from.

Those are the only resolution and hardware disclosures in the post. Pika does not state the output resolution for Live Staging or Live Actor Direction, does not publish a prompt set or harness, and gives no figure for how many sessions a node can serve at once. A latency number on a single H100 and a time-to-first-frame number on eight B300s describe different things on different machines, so the three multiples do not stack into a single claim about how fast Enact runs.

The control stack is synthetic data and a sliding cache

The engineering disclosure is the more substantial part of the post. Real footage does not come with frame-aligned cursor trajectories, so Pika built a synthetic pipeline that animates a 3D stand-in along a drag path with simulated weight and momentum, renders it photorealistically with Pika's own video models, and records every mouse position and button state. At inference, three frame-aligned signals feed the base model through lightweight zero-initialized modules: mouse state sampled at 24 fps, a heatmap marking the grabbed point, and the untouched source video to hold the background in place.

Streaming comes from causal chunking, a sliding KV cache that keeps the prompt, the source video, the first chunk and the four most recent chunks, re-based positional embeddings, and self-forcing distribution matching distillation that cuts the sampler from roughly twenty steps to a few. A component Pika calls Real-Time Director injects periodic hold beats to stop an idle scene from drifting away from its staging.

Open base model, closed everything else

H3 Flex itself shipped as open weights capped at 768p, but Pika has not said that Enact's real-time control models, synthetic data pipeline, orchestration layer or streaming implementation will follow. The post lists no waitlist, no sign-up, no pricing and no demo build, only an email route for enterprise conversations. There is also no stated session length, no drift measurement over time and no example of a long continuous take, which is the one thing a live-direction tool would need to demonstrate. Pika's published API rates for its existing video endpoints start at $0.01 per output second for text or image input and $0.03 per second for reference-guided video; what Enact would cost is unstated. Pika has not given a date for wider access.

Sources: Pika, Pika on X.

Cite this

Free to cite and reuse with a link back. Data is updated as new runs and prices come in, so include the date.

Genativity. (2026). Pika's Enact preview publishes three real-time speed figures measured three different ways. Retrieved November 20, 2026, from https://genativity.com/news/pika-enact-three-real-time-figures-two-hardware-setups
<a href="https://genativity.com/news/pika-enact-three-real-time-figures-two-hardware-setups">Pika's Enact preview publishes three real-time speed figures measured three different ways</a> (Genativity)

Daniel Ochoa

Daniel Ochoa: Covers model launches, shutdowns and pricing changes as they happen. Reads deprecation notices for a living so you do not have to.