Runway has shown research into real-time video generation, in which users stream video as they describe it instead of entering a prompt and waiting for a finished clip. The company says users repeatedly report losing the most time generating and revising videos, and it wants to minimize the time to the first frame. For businesses that build on generative video, this changes the economics of the tool rather than just its speed.
What Runway showed and when
The approach was first discussed in March with Runway Characters. It relies on GWM-1, the company's first General World Model, introduced in December 2025. GWM-1 builds on Gen-4.5, generates video frame by frame, and accepts camera movements, robot commands, or audio as controls. A few weeks ago Runway also showed Solaris, a system that uses Gen-4.5 to generate user interfaces frame by frame and responds to clicks or voice input. No availability timeline has been announced.
Today's video models work in separate steps: a prompt goes in, a few seconds or minutes pass, and a finished video comes out. If the result is wrong, the user starts over. Runway's answer is to keep the model running while the user steers, so that most of the session time goes into directing the video rather than waiting. The company frames this as closing the gap between an idea and its execution. The same mechanism underlies the Solaris interface work, where each frame is produced in response to input rather than pre-rendered.
Runway also points to lower costs. Faster models consume less GPU time, which makes them more cost-efficient, and the company argues that the cost per output at a given quality level determines which applications make economic sense. Instant generation would lower that threshold, bringing previously unprofitable applications into range. According to Runway, real-time generation shifts the compute load from training to use: the model must produce each frame fast enough to keep up with playback while running on hardware shared by several sessions at once.
What this means for business
For companies that use AI video, the practical consequence is a different cost structure rather than a new creative option. When the price per finished clip falls, tasks that were too expensive to automate become viable: routine product footage, training material, or scenario variants that are currently cut for budget reasons. A small team is likely to notice this first in short, repetitive formats, where the volume of clips matters more than their polish. A large company with its own GPU capacity will weigh the same change against infrastructure it already owns, and its calculation will depend on how many sessions share one machine.
Several things remain unverified. Runway has not announced when the technology will be available, and the research preview shown with Nvidia at the chipmaker's GTC conference, running on the Vera Rubin platform, is designed to deliver the first frame in under 100 milliseconds but has no stated release date either. A video model builds each frame on the previous one, so small visual errors can compound into major distortions over time; Runway calls this the central problem with LLM-based approaches and trains the model on its own outputs, not only error-free inputs, so it learns to correct deviations instead of amplifying them. Startup Decart used a similar method for its real-time model MirageLSD, deliberately exposing it to flawed or distorted images during training. Before committing, buyers should ask how long a session can stay coherent, how many sessions share one GPU, and what quality level the cost figures refer to.
The sign to watch is a published availability date and a stated cost per output at a defined quality level. Until Runway provides both, real-time generation remains a research direction rather than a purchasable capability, and budgets should be planned on the current step-by-step models. If a date appears together with per-output pricing, the threshold Runway describes becomes something a business can actually calculate against its own production volume.
