Direct a long take with GPT-6 Astra and Seedance 2.5

That is one room of the take, unedited, from a real session. The clip above is the generated segment itself. Multiply it by the rooms in your plan and you have the raw material for the walk-through; joining the segments is ordinary ffmpeg work the agent does locally for free.

The workflow behind the seven-room shot

In that demo, GPT-6 Astra mapped the seven rooms and kept the blocking inside the viewport while Seedance 2.5 rendered the shot. Astra was released on September 3, 2026 around long-running agent tasks and computer use; spatial planning and shot logic sit squarely inside that, and since Codex CLI 0.153.4 it is the bundled default model. What the model does not do is paint pixels — that stays with a video model.

From a Codex session the split looks like this. You describe the location and the path — seven rooms, one continuous move, where the camera turns. The agent writes the shot plan: a prompt per room that carries the light, the doorway, and the direction of travel so the segments cut together. Each prompt becomes one priced Seedance 2.5 run through Actionway's video generation capability; the agent waits on the jobs, downloads the clips, and concatenates them with ffmpeg. The plan is the agent's work and costs nothing; only the renders are paid.

Pick a model, pay per second

Long takes reward consistency, so pick one model for every segment. Seedance 2.5 (from $0.20 per second (480p)) is the model in the original workflow and accepts clips from 4 to 30 seconds; Seedance 2.0 (from $0.14 per second (480p)) reaches 4K when the final delivery needs it; Kling 3 (from $0.20 per second (720p)) is a second look for the same plan. The agent lists these neutrally with prices and you choose — say "you pick" and it takes the best-quality tier. Draft the whole path at 480p first: at draft prices a seven-room pass is cheap enough to reshoot a room whose blocking drifted, and the premium resolution is spent once, on the final assembly.

Render time is the other budget. The 2.5 model took about eight minutes for this five-second segment, so a seven-room plan is a job the agent waits on rather than a live preview; queue the rooms, let the agent collect them, and review the assembled take. Every segment is one catalog-priced run, and there is no subscription — a balance you top up once covers every capability in the catalog.

Keeping seven rooms consistent

Continuity is the hard part, and it is mostly a prompt problem. Reuse the same opening clause in every segment — the light, the time of day, the lens — and change only the room and the doorway. End each prompt with the direction the camera leaves the frame and start the next with the direction it enters, so the cuts have somewhere to hide. Keep the resolution and the model the same across the whole take; mixing Seedance 2.5 and Kling 3 segments gives you two colour palettes in one corridor.

Review the segments before joining them. The agent can pull the first and last frame of every clip into a contact sheet locally, which is the quickest way to spot a doorway that opens the wrong way or a room that came back at night. Reshoot only that room: at $0.20 per second (480p) for a 480p draft, a second attempt at one segment costs far less than re-rendering seven. When the assembled 480p take plays through, run the same seven prompts once more at the delivery resolution, and join those.

Set it up once

npm install --global @actionway/cli@latest
actionway init
mkdir -p "$HOME/.codex/skills"
cp -R "$HOME/.actionway/skill/actionway" "$HOME/.codex/skills/actionway"

You need Node.js 20 or newer and an Actionway account; after sign-in the agent reports your wallet balance and, if it is empty, links you straight to top-up. Picking a priced option is the approval — the run goes ahead at the shown price, and only runs the platform flags for extra confirmation pause first. The get started page lists every route per client, including the Codex plugin.