Summary
- Runway released a 54-second demo on September 11. Between a single product photo and a finished commercial, GPT-6 Astra opened and drove Runway and Blender itself.
- The job runs in four stages: a white clay style frame built in Runway, camera paths and animation assembled in Blender with Eevee, a final render handled by Seedance 2.5 instead of a local GPU, and a repeat pass once the instructions change.
- Installing the Runway plugin in the ChatGPT app is enough to use it, and usage is drawn from an existing Runway plan with no separate API key. Model selection criteria, per-model pricing and render speed figures were not published.
Between a single product photo and a finished commercial, the only thing the human did was give instructions in a chat window. This is the 54-second demo published on September 11 by Runway, which builds AI video generation tools, and in the middle of it ChatGPT opened and drove both Runway and Blender, the 3D production program, on its own. The entire order of operations for making a 3D commercial has moved inside the chat window.
Until now this work required a person to handle Blender directly. Placing objects, plotting the path a camera travels and setting lights are skills that take months just to learn. In this demo the seat was taken by GPT-6 Astra, OpenAI's reasoning model, and the person described the picture they wanted in words.
In the 54-second video METAL reviewed, the job runs in four stages. The first is the style frame, a reference shot that fixes the color, texture and lighting of the finished piece in advance. Astra passed the product photo to Runway and had it build that shot in a white clay texture. Strip the color out and leave only form and light, and the structure holds no matter what style gets applied later.
In the second stage Astra opens Blender. Inside the 3D program it places objects, plots the camera's path and then renders the frames with Eevee. Eevee is the fast render engine built into Blender, which makes it good for checking results on the spot. It is the place where you choose speed over finish.
In the third stage the work returns to Runway. The Blender frames go back with a set of style reference images, and the final render is handled not by a local GPU but by a video model called Seedance 2.5. The Runway MCP product page lists Veo 3.1, GPT Image 2, Kling 3.0, Seedance 2.5 and Nano Banana Pro alongside Runway's own Gen-4.5, Gen-4 Image and Gen-4 Turbo, so a model suited to the job can be called.
The fourth stage is repetition. The video explains that asking for a different camera path or a more complex movement runs the same process again. Much as a director gives separate notes to a cinematographer and a production designer, the person never enters each program and steers from one seat, in words.
Two things change in practice. Someone who cannot operate Blender can now try a 3D look, and there is less reason to sit holding a machine while a final render finishes. The video too leads with the ability to try many versions and styles quickly rather than with time saved.
There is a small joke buried in the video. The presenter says he is stepping out for a coffee, then immediately adds that the figure on screen is an AI version of himself and that the real him watched the whole thing happen live. In a demo from a company that makes production tools, the fact that even the host is generated gets handled in passing.
To try it, install the Runway plugin in the ChatGPT app. The company's product page reads "Now in the ChatGPT app directory," and it explains that once installed you call Runway in the middle of a conversation. The underlying connection standard is Runway MCP, released on May 27. In that announcement the company said, "Today we're launching Runway MCP (Model Context Protocol), an MCP server that connects Runway to all of the agents and coding tools you already work in."
Where the money comes from is stated as well. Runway's announcement reads, "No separate API key is required - generations are tied to your existing Runway plan." Whether the call comes from ChatGPT or from Cursor, usage leaves the Runway account. Cristobal Valenzuela, co-founder and co-chief executive, wrote only "Use Runway inside ChatGPT" on his X account on June 15.
METAL has reported on Higgsfield building a Seedance 2.5 video from Blender 3D blocking, and this demo walks essentially the same path. The one difference is whether the blocking is assembled by a person's hands or by an agent.
Parts of this demo are hard to take at face value. The video is marketing made by the company, and no figure was given for how much faster it is than a local GPU. Which model is attached to which job, and how pricing differs between them, was not disclosed either, so the result arrives first and the cost is confirmed afterward.
METAL has also covered Runway's release of a playable world model with sound, where what the company led with was the performance of the model itself. This time what it put forward is not a model but the place from which a model gets called. Making a good tool and being called well have become two different problems.
In 3D commercial production, the point where a human hand touches the work is moving from the program to the instruction. Someone who has never opened Blender receiving frames made in Blender already happened in this demo. A creator's craft is shifting from dexterity with tools toward the ability to say precisely what they want.
Comments