LATEST
AI Tools

Pika Enact: Direct AI Video Live, Like a Film Set

Updated November 20, 2026
14 min read
Editor's Note: This article was human-reviewed for accuracy
Behind the scenes on a film set, with a director and crew working around a camera

Quick Answer & Key Takeaways

  • • Pika Enact is a research preview that turns AI video generation into live directing: you can move actors, give the cast spoken notes, and steer the camera while the scene keeps playing.
  • • It is built on Pika H3 Flex, it is not available to the public yet, and Pika has published no price for it.
  • • The paid plans underneath it start at $10 a month, and a five-second H3 Flex clip costs 10 credits on that plan.

AI video has worked the same way since the first text-to-video tools appeared: you write a prompt, wait for the render, and hope the result matches what you pictured. If it does not, you write another prompt and wait again. Pika Enact, announced on November 18, 2026, tries to break that loop. It turns generation into a live session where you move the actors, talk to the cast, and steer the camera while the scene is still playing.

Enact is not a product you can sign up for. It is a research preview, built on Pika's H3 Flex model, and Pika says it is unavailable for consumer use. That matters, because the demos are striking and easy to mistake for a launch. This article covers what Enact does, the numbers Pika published for it, what the platform underneath it costs, and the limits Pika itself has stated.

What Pika Enact actually is

Pika describes Enact as an experimental live-streaming environment designed to mimic directing on a film set. It is built on Pika H3 Flex, the company's post-trained MiniMax H3 model, and it generates video and sound in sub-second chunks while carrying scene context forward. New direction goes into the generation as it happens, rather than into the next render.

Three modes sit on top of that streaming setup. Live Staging lets you drag actors and objects around a playing scene with your mouse. Live Actor Direction lets you speak to the people in the scene and give them notes. Live Camera Control lets you steer a virtual camera with your mouse while the scene keeps generating. Pika calls this an early look, not a release, and anyone interested in trying it is directed to its enterprise team at [email protected].

The distinction matters. Enact is a system on top of the existing H3 Flex model. It adds interaction, streaming, and an agent layer, not a new foundation model, and nothing in the announcement gives a date for public access.

The three ways you direct a scene

Live Staging: drag the actors while the scene plays

Blocking a scene, deciding where each actor stands and when they move, is ordinary film work. In prompt-based AI video it is guesswork: a prompt can suggest "a bit to the left" and nothing more. With Live Staging, you click an actor or an object and drag it to the exact spot while the video keeps playing.

Pika says the subject responds to being moved rather than sliding like a cut-out. As you drag, its pose, lighting, and shadows adapt to the surroundings, and the person shifts weight and adjusts posture mid-air before settling where you let go. The result, according to Pika, has no cut-out edges, no hole where the subject stood, and no compositing pass to clean up afterwards.

The speed figures Pika published for this mode are specific to its own test setup. On a single server, the model generates the video 1.8 times faster than it plays. In the low-latency setting, the picture responds in under 0.4 seconds on an NVIDIA H100, decoding included. The model streams short causal chunks of 0.7 seconds in the standard setting and 0.17 seconds in the low-latency setting.

Training a model to do this needed data that real footage cannot supply: strictly consistent before-and-after video pairs, synchronised with frame-level cursor movements. Pika built a physics-driven synthetic data engine for it. Starting from an untouched video, the engine animates a 3D stand-in along a drag path to simulate weight and momentum, then renders photorealistic target clips while logging mouse positions and button states.

Live Actor Direction: talk to the cast, then call Action

Once a scene is staged, Enact lets you direct the people in it. You type or speak a note, such as "move closer to her" or "make it a kitchen at night", and the actors play it out while the video keeps running. They answer you in character, in their own voices, and Pika says they remember your notes, so you never have to repeat one.

The workflow Pika shows is rehearsal first, then a clean take. You work through a performance line by line, giving notes and watching the actors try them. When it feels right, you call "Action", and everything rehearsed becomes one fresh take, shot from the first frame. Every line of the rehearsal is logged verbatim, and the script model rewrites that exchange into one uninterrupted performance.

Underneath, this mode is an agentic model that thinks and performs at the same time. A script language model reads your notes with the opening frame and session memory, then turns them into structured beats covering action, dialogue, emotion, gaze, and lighting. A real-time video model does the acting, and when no instruction is queued the scene holds naturally with subtle breathing and blinking. Pika says the picture and sound generate 1.75 times faster than they play in this mode.

One detail from Pika chief executive Demi Guo, speaking to The Neuron, is worth keeping in mind: the actors' memory is session context managed by software, not evidence that the characters understand anything. The system tracks who got which note, so a direction meant for one actor lands on that actor and not the other, and it carries that context through the scene.

Live Camera Control: move the lens while the take runs

The third mode separates camera work from performance. You steer a virtual camera with your mouse while a scene is continuously generating, panning, dollying in for a close-up, or orbiting around the subjects. An appearance image anchors the cast and set, while a lightweight 3D motion proxy, such as an untextured Blender viewport, guides the camera path. No existing video is needed: the system generates audio and video from scratch.

Pika's published benchmark for this mode ran on an 8x NVIDIA B300 node at 640x480 and 24 frames per second. Initial frames render in 0.51 seconds, and subsequent chunks generate at 1.50 times real-time speed. That resolution is low, and it is a real constraint on what the demos prove: a 640x480 stream is a long way from the 4K finish Pika advertises for its non-real-time model.

The point of separating the modes, as Pika's creative lead Matan Cohen-Grumi put it in the same Neuron conversation, is that you can lock in a performance you like and then try different camera movements without asking the cast to recreate the moment. Performance and camera become two decisions instead of one gamble.

The engine underneath: Pika H3 Flex

Enact runs on Pika H3 Flex, which Pika introduced on October 29, 2026. Flex is a post-trained version of the MiniMax H3 model, built for creative iteration. Pika says it can generate a 15-second clip with synchronised audio at 768p in as little as seven seconds of transformer computation, after post-training cut generation from 50 denoising steps to six. That seven-second figure measures transformer computation in a controlled benchmark, not total end-to-end API latency, and Pika says so in its own footnote.

Flex accepts up to 12 reference files, combining images, video, and audio in one generation, and it can export five separate audio stems: Dialogue, Score, SFX, Foley, and Ambience. Clips are drafted at 768p and can be upscaled to 1080p, 2K, or 4K. At launch, Pika said Flex ranked first among open-weights models and seventh overall on DesignArena's text-to-video leaderboard, a snapshot position rather than a standing result.

Pika released Flex's generator adapters and inference code as open weights, with downloads on Hugging Face and setup instructions on GitHub. The release covers text-to-video, first- and last-frame generation, and multimodal reference-to-video at 480p and 768p. The upsampler, stem separation, and accelerated serving stack are Pika-hosted features, so local results depend on your hardware.

What it costs

Enact itself has no published price. It is a research preview with no consumer access, so there is nothing to buy and no plan to attach it to. Pika directs interested parties to its enterprise team, which means pricing, if it ever gets one, will start as a conversation rather than a rate card. What does have a price is the Pika platform underneath it, and that is the honest way to cost the tools you can use today.

Pika's plans, read from its pricing page on November 19, 2026, are Free at $0, Starter at $10 a month or $8 a month billed annually, Creator at $35 a month or $28 a month billed annually, and Fancy from $95 a month or from $76 a month billed annually, up to $880 a month. Every plan includes all models. The Free plan has no watermark but no commercial licence; the paid plans include a commercial licence. Plan credits refresh each billing period and do not roll over. Top-up credits carry over and expire 365 days after purchase.

PlanUSD / monthGBP (approx.)CAD (approx.)Monthly creditsSuited to
Free$0£0CA$00 (credit packs only)Trying the platform
Starter$10 ($8 annual)£7.60 (£6.08)CA$14.30 (CA$11.44)600Occasional clips
Creator$35 ($28 annual)£26.60 (£21.28)CA$50.05 (CA$40.04)2,600Regular social content
Fancy$95–$880 ($76–$704 annual)£72.20–£668.80 (£57.76–£535.04)CA$135.85–CA$1,258.40 (CA$108.68–CA$1,006.72)8,550+Heavy production use

Currency conversions use November 20, 2026 market rates of 1 USD = 0.76 GBP and 1 USD = 1.43 CAD, and are approximate. Annual plans are paid upfront.

Credits are where the price turns into a per-clip number. A five-second Pika H3 Flex clip at 768p costs 10 credits, so the Starter plan's 600 credits buy 60 such clips a month and Creator's 2,600 buy 260. On Starter, that works out to about $0.17 per H3 Flex clip by simple division of the $10 monthly price by 600 credits, our arithmetic, not Pika's. Other models cost far more: a five-second Seedance 2.5 clip at 720p costs 122 credits, which Pika's own pricing FAQ puts at about $2.03 on Starter. A five-second Kling 4.0 clip at 1080p costs 118 credits.

Credit packs are a separate one-off purchase: $25 for 1,500 credits, $50 for 3,000, $200 for 12,000, $500 for 30,000, and $1,000 for 66,000 credits, with a 10 percent bonus on the largest pack. Top-ups work only on Pika Create and Pika MCP, not with the API, the iOS app, or the old Pika site.

For developers, Pika's technical report lists starting rates of $0.01 per output second for video made from text or an image, and $0.03 per second for video guided by references, as reported by The Rundown. Those API rates describe H3 Flex generation, not the live directing environment, and Enact's own pricing remains unknown.

Enact vs the alternatives

Enact is not competing with other video generators on clip quality. It is competing on the workflow around the clip. The table compares it with the closest options, using each product's own published specifications.

ToolHow you direct itPricePublic access
Pika EnactLive: drag actors, talk to cast, steer camera while it playsNone publishedNo, research preview
Pika H3 Flex (standard)Prompt plus up to 12 references, then re-rollFrom $10/month; open weights free to runYes
Kling 4.0Prompt plus 10 keyframes; 30-second clipsCredit-based, via Pika or KlingYes
Standard text-to-video toolsPrompt, wait, re-rollCredit or subscription, variesYes

Our Kling 4.0 review covers the 30-second, 10-keyframe model in detail, and our Kling Image 4.0 review covers its image counterpart. For the wider field, our FLUX 3 Max review covers a 12-second video model with a very different price structure.

Google is moving in a similar direction. Its Genie 4 world model began rolling out in Project Genie on November 18, the same day as Enact, giving Google AI Ultra subscribers a world they can act inside rather than a clip they watch. The two systems are different in kind, a world model against a directing environment, but both point at the same shift: generation that responds while you use it, not only before you start.

What it cannot do yet

The limits are as informative as the demos, and most of them come from Pika's own announcement or its chief executive.

First, you cannot try it. Enact is a research preview, unavailable for consumer use, with no published release date. The only route in is Pika's enterprise team. H3 Flex being available, including as open weights, is not the same as Enact being available, a distinction Pika's chief executive made explicitly.

Second, the live experience trades quality for speed. Demi Guo told The Neuron that the responsive, real-time model is slightly lower in image quality than Pika's non-real-time version. That is the cost of generating faster than playback, and it is why the camera benchmark runs at 640x480 rather than at the 4K finish Pika advertises elsewhere.

Third, every speed figure in this article is Pika's own. The 1.8 times, 1.75 times, and 1.50 times real-time figures were measured on Pika's hardware, in Pika's test setups, with no independent test published yet. The 0.4-second staging response was measured on an H100, and the camera benchmark ran on an 8x B300 node, hardware few creators have at home. Treat these numbers as a claim about what the architecture can do under lab conditions, not a promise about your machine.

Fourth, the actors' memory is a session mechanism, not understanding. The system logs your notes and carries them through the scene, which is useful for a rehearsal. It is not evidence that the characters know anything, and writing about it as though they do would mislead readers about what the tool is.

Common mistakes

Do not treat Enact as a product you can sign up for today. It is a research preview, and the sign-up page for Pika sells you the standard platform, not Enact access.

Do not confuse H3 Flex availability with Enact access. The model is open weights and already in the Pika product. The live directing layer on top of it is not.

Do not quote the 1.8 times figure as a general speed claim. It describes Live Staging on a single server in Pika's own setup, and the other two modes have their own figures on their own hardware.

Do not expect 4K live directing. The camera benchmark Pika published runs at 640x480, and the real-time model gives up some image quality for responsiveness. The 4K finish belongs to the non-real-time workflow.

Do not read the actors' memory as understanding. It is logged session context that keeps a note attached to the right actor. That is a practical feature, and it does not need a mystical description.

Frequently asked questions

Is Pika Enact available yet?

No. Enact is a research preview announced on November 18, 2026, and Pika says it is unavailable for consumer use. The only route Pika offers is contacting its enterprise team at [email protected]. No public release date has been announced.

How much does Pika Enact cost?

Pika has published no price for Enact, because there is nothing to buy yet. The Pika platform underneath it starts at $10 a month for the Starter plan, or $8 a month billed annually, with 600 credits a month. A five-second H3 Flex clip at 768p costs 10 credits on that plan.

What is Pika H3 Flex?

H3 Flex is Pika's post-trained version of the MiniMax H3 model, introduced on October 29, 2026. It generates 15-second clips with synchronised audio at 768p, accepts up to 12 reference files, exports five separate audio stems, and finishes at up to 4K. Its generator adapters and inference code are released as open weights.

Can I run H3 Flex on my own hardware?

Yes, within limits. Pika released the generator adapters and inference code, with weights on Hugging Face and setup instructions on GitHub. The release covers text-to-video, first- and last-frame generation, and multimodal reference-to-video at 480p and 768p. The upsampler, stem separation, and accelerated serving stack are Pika-hosted features, so local results depend on your hardware.

How is Enact different from normal AI video generation?

Normal generation takes a prompt, renders a finished clip, and makes you start over to change anything. Enact streams the generation in sub-second chunks and lets you change it while it plays: drag an actor, give the cast a spoken note, or move the camera. The performance and the camera become decisions you make during the take, not before it.

Will Enact come to Pika inside ChatGPT?

Nothing announced connects Enact to the Pika extension inside ChatGPT, which arrived on October 16, 2026. Enact is a separate research preview with enterprise-only access, and Pika has said nothing about bringing the live directing modes to any chat product.

Sources

Pika Enact is a research preview, not a product, and the honest read is that its most useful part today is the model underneath it. H3 Flex is already available, already open weights, and already priced at 10 credits for a five-second clip. Enact shows where that model could go once directing happens live instead of in advance. Until Pika opens it up, the thing to watch is whether the quality tradeoff narrows and whether the enterprise-only preview ever becomes a plan you can buy. If you make AI video, try H3 Flex on the Starter plan: that is the part of this story you can test today.

14 min read
38 views
edit_note Written by

Ratul Ahamed

Founder & Editor-in-Chief

Ratul Ahamed is the founder and editor of AI Core Updates, an independent site covering AI tools, models and pricing. He is a web developer based in Bangladesh and also runs All In One Tools, a free collection of browser-based utilities. Contact: [email protected]

Read all stories by Ratul Ahamed arrow_forward
verified_user Human-Reviewed Editorial

forum Comments (0)

Leave a Reply

chat_bubble_outline

No comments yet. Be the first to join the conversation!

Weekly AI Digest

Get our weekly AI tools digest

One email a week analyzing new AI tools, model updates, and practical workflows. Tested and summarized with zero fluff. Unsubscribe any time.

bolt
Hands-on AI Tool Reviews
Pricing math, limits, and alternatives.
verified_user
Strict Anti-Spam Policy
1-click unsubscribe at any time.
mark_email_unread

Subscribe to Digest

Sent weekly • No spam

person_outline
mail_outline

No spam ever. You can unsubscribe with a single click at any time.

cookie

Cookie Settings

We use cookies to enhance your browsing experience, serve personalized content, and analyze our traffic. By clicking "Accept All", you consent to our use of cookies. Learn more