GPT-6 Astra Made a Full YouTube Video From One Prompt
GPT-6 Astra researched, scripted, voiced, and edited a complete YouTube video from a single prompt. Here's how the pipeline actually worked.
What happened
A creator gave GPT-6 Astra a single open-ended prompt: take an idea to a finished YouTube video about Astra’s own release, using his voice clone and avatar, and produce something ready to publish, not a draft. Astra came back with a fully edited video: it researched other people’s Astra projects, wrote a script, generated narration through a cloned voice, drove an AI avatar, cut the footage together with timed music and sound effects, and rendered a final file. The creator behind the experiment (Nate Herk) reported the whole job took about 50 minutes and would have cost roughly $60 in API billing.
This matters less as a novelty and more as a preview of what “computer use” agents look like when pointed at a real, multi-step creative workflow instead of a toy task.
TL;DR
- A single open-ended prompt (“take me from idea to finished YouTube video”) was enough for GPT-6 Astra to plan, research, script, voice, edit, and render a complete video without step-by-step direction.
- Computer use let Astra browse and verify sources itself, opening other creators’ original social posts to check what was actually shown before referencing their work in the script.
- The video used existing tools, not custom models: HeyGen Avatar V5 for the talking presenter, an ElevenLabs voice clone for narration, and a tool called HyperFrames for the actual timeline edit.
- Astra split narration into segments, sent audio to the avatar tool, and organized results rather than generating one long clip, which is closer to how a human editor would structure the work.
- The finished render was checked against the script, with Astra transcribing its own output audio and comparing it back to catch clipped words or covered dialogue.
- The whole job took about 50 minutes and was estimated at around $60 if billed through the API, though the creator ran it in a faster, pricier mode.
- OpenAI released GPT-6 Astra on September 3rd, with computer use and longer task duration as the two headline capabilities, and access rolling out in stages rather than to everyone at once.
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
How did Astra actually build the video?
The process broke into distinct phases that mirror a real production pipeline rather than a single generation step.
First, research. Astra used computer use to open the original social media posts referenced in the prompt’s context, capturing pages and checking what was actually on screen rather than relying on secondhand descriptions. That included a Manhattan-scale environment built in Unreal Engine, a Call of Duty style shooter played for roughly two hours, a Minecraft style build described as one shot, an old train drawing converted into more than 3,000 editable objects in Blender, a real estate walkthrough built from listing photos, and a motion graphics piece with a reported 14-minute generation time. Astra treated these as source material to verify, not just quote, flagging in its narration where a creator had noted inaccuracies in their own results.
Second, scripting. Astra wrote narration tying these examples together around the theme of “what people are building with Astra,” structured as a tour through project files rather than a generic feature list.
Third, voice and presenter. Astra split the script into short segments and sent each to HeyGen Avatar V5, paired with an ElevenLabs voice clone of the creator, so the final video shows an avatar speaking in the creator’s own voice while narrating that it is not actually him.
Fourth, editing. The assembly happened inside HyperFrames, described as a timeline tool where Astra could control the timing of camera moves, individual words, clips, and transitions, then build sound design (music under longer movements, a click on interface changes, brief pauses before new shots land) around those cuts.
Fifth, verification. After rendering, Astra checked frames from the exported file and transcribed the finished audio to compare it against the original script, a step aimed at catching mismatches like a clipped title card or a sound effect stepping on dialogue.
Why does computer use matter here?
Computer use is the capability that let Astra act like an editor with a browser and a file system instead of a model that only outputs text. In this workflow, that meant three concrete things: opening real external webpages to confirm what other creators actually posted, navigating the creator’s own project folders to find an existing avatar setup, voice clone, and connected tools already in the workspace, and operating the HyperFrames editing tool directly to place clips and transitions on a timeline.
That last part is the one worth sitting with. Rather than generating a rough cut and handing it to a human to finish, Astra performed the fine-grained timeline work itself, which is the part of video production that usually eats the most hours.
Was the prompt actually simple?
One coffee. One working app.
You bring the idea. Remy manages the project.
The prompt was short but not blank. The creator specified that the video should use his existing voice clone and avatar, that it should open by clarifying the presenter isn’t really him and that Astra orchestrated the whole thing, and that it should be a finished, postable version rather than a draft or proof of concept. He also added a line asking Astra to verify its work rather than hand back a first pass.
Beyond that framing, the creative decisions (which examples to feature, how to structure the script, where to place music cues, how to pace the sound design) were left to the model. That’s a meaningfully different kind of prompt than “write me a script” or “generate a clip.” It’s closer to a creative brief handed to a freelance editor than a generation request handed to a text model.
Is this actually production quality?
The honest answer is that it depends on your bar, and the evidence here is a single self-reported example, not an independent benchmark. What’s notable is the process discipline built into the run: Astra cross-checked source posts before referencing them, flagged another creator’s own admission that his walkthrough had inaccuracies, and ran a post-render transcription pass specifically to catch audio and timing errors. That’s a workflow that treats accuracy and verification as steps, not an afterthought.
What the transcript doesn’t establish is how much manual cleanup, if any, happened after this render, or how this output compares to a professional editor working the same brief. The creator did note he ran the job in a faster, more expensive mode of the model and had already burned through multiple usage resets experimenting with it, which suggests iteration was part of getting a result worth showing.
What does this cost, in practice?
The creator estimated the job at around $60 if billed through the API at standard rates, noting the actual run used a faster mode that costs more. Total time from prompt to finished render was about 50 minutes. Those figures come from one experiment on one account tier, so they’re a data point, not a stable price sheet, and OpenAI’s staged rollout of Astra access means availability and limits will likely shift.
Frequently Asked Questions
What is GPT-6 Astra?
GPT-6 Astra is a model OpenAI released on September 3rd, built around two headline capabilities: computer use (the ability to operate a browser, files, and connected tools directly) and support for longer, multi-step tasks than earlier models handled well. Access has been rolling out in stages rather than all at once.
What tools did Astra use to make the video?
Astra used HeyGen Avatar V5 to generate the on-screen presenter, an ElevenLabs voice clone for narration, and a timeline editing tool called HyperFrames to assemble clips, control transition timing, and build sound design. Astra did not build these tools itself; it operated existing accounts and integrations already set up in the creator’s workspace.
Did a human edit or clean up the final video?
The transcript describes Astra performing research, scripting, voicing, editing, rendering, and a post-render verification pass (transcribing the output and comparing it to the script) without describing a separate manual editing pass afterward. The creator supplied the initial brief and feedback but described his role as directing rather than producing.
How much did it cost and how long did it take?
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
The creator reported the full job took about 50 minutes and estimated it would cost roughly $60 if billed through the standard API, though he ran it in a faster mode that increases cost. He also noted burning through multiple usage resets while experimenting with the model.
Can anyone replicate this exact workflow today?
Not necessarily on the same terms. Access to GPT-6 Astra is rolling out in stages, and this specific setup relied on tools and accounts (a voice clone, an avatar license, a connected editing workspace) already configured before the prompt was given, so the underlying capability is real but the reproduction requires the same supporting infrastructure.
