Enjoying this issue?
Get tomorrow's AI & engineering digest in your inbox — hand-picked, summarized, and always spam-free.
TLDR
GPT-6 Astra can now generate a complete YouTube video from a single prompt, including script, footage, voiceover, and editing. The demo showed a professional-quality result using the user's voice clone and avatar, costing roughly $60 in API fees for 50 minutes of generation.
Key points
OpenAI released GPT-6 Astra on September 3rd with computer use and longer tasks.
Astra created a full video from one prompt, including script, B-roll, editing, and sound.
The presenter used Nate Herk's voice clone and Henan Avatar V5 for narration.
The generation took about 50 minutes and would cost approximately $60 in API billing.
Astra can build 3D worlds and games, as shown by examples like Matt Schumer's Manhattan world and Riley Brown's shooter.
Tools mentioned
Techniques
- computer use
- longer tasks
- editing in hyperframes
- generation with verification
Stop scrolling. Start reading smarter.
Receive the day's most important AI & engineering updates in one concise email. No spam.
Transcript (captions)
You guys are not going to believe how ridiculous this is. I gave Astra this one prompt and this is what I got back. I'm Astra 6. You're looking at Nate Herk's avatar speaking with his voice
clone. I made this video. I found the footage, captured the posts, wrote the script, and built the edit. The music and sound effects. I handled those, too. Nate asked for a video about my release
and what people are building with me. So, let's open the files. There's a city, a shooter, a Minecraft style game, and the production you're watching right now. Matt Schumer shared this Manhattan
world built with Astra in Unreal Engine. [music] Watch the camera drop from the skyline to street level. Buildings, traffic, storefronts. His post says the project ran over a week with Astra
working street by street. He also says finding a setup that could run that long took work. This is the kind of assignment I can keep returning to. Inspect a scene, make changes, and check
the result. Riley Brown took that loop into a game. Riley posted this Call of Duty style shooter and said he played for about 2 hours [music] asking GPT6 for changes between matches. You're
looking at his actual gameplay over here. Flavio shared a Minecraft style build he described as one shot. The clip shows movement, an inventory, crafting, and breaking blocks. Those are creator
reports, and I've linked their original posts. [music] I like that you can see people interacting with the results. Now, watch what happens when the output needs to stay editable. Tom Kerchie gave
Astra an old train drawing. He reports more than 3,000 editable objects in Blender. Watch him select one piece. That's a file someone can open and keep working on. Yunf used listing photos to
build this house walkthrough. He flagged wrong details himself, so you'd still need to compare it with the real place. I work with that same discipline on a video. Keep the source, inspect the
output, and fix what doesn't match. Daniel th shared this motion video and reported 14 minutes of generation plus two changes. Look at the way the interface moves through the frame. I
used that as a visual reference for this edit. OpenAI released GPT6 Astra on September 3rd with computer use and longer tasks as major themes. Access is rolling out in stages for this job.
Computer use meant opening the original posts, capturing the pages, and checking what was actually on screen. Then I worked through Nate's project folders. I found his avatar setup, his voice clone,
and the tools already connected to this workspace. I split the narration into short segments, sent the audio to Hunen Avatar V5, and organized the results beside the original demos. This blue
background is Nate's own image. I brought that in, too. The edit lives in hyperframes where I can control the timing of a camera move, a word, a clip, or a transition. I built the sound
around those cuts. Music carries the longer moves. A click lands on an interface change. A short break gives the next shot room to hit. Then, I rendered the video and checked frames
from the exported file. I also transcribed the finished audio and compared it with the script. If a title gets clipped or a sound covers a word, I can go back to the project and fix it.
That's how I get from an idea to a file you can use. You've just watched the result. The presenter is Hen Avatar V5. The voice is Nate's 11 Labs clone. I directed those tools and built the
Hyperframes edit around real source material. Nate gave me the assignment and feedback. I carried out the production. You can open the video, the subtitles, and the source list, and you
can judge the work for yourself. Give me a project with a clear finish line. What would you have me build? I mean, that is absolutely crazy. So, let me show you guys the prompt and then
tell you how much that actually cost. So, I said, I need you to take me from idea to finished YouTube video. I want you to create me an engaging and professional video about GBT6 Astra. You
need to use Nate Herk's voice clone and Hen avatar. So, I did give it that context. I said, "Start the video off by telling everyone that this is not Nate and that you orchestrated the entire
video. You planned it out. You researched it. You grabbed all the B-roll. You grabbed the screenshots, you edited everything, and you produced a finished video, which is what they're
watching right now. I want this to feel professional, like it was edited by a human using hyperframes. Do not give me a draft or a PC. Give me a finished working version that I can post on my
YouTube channel. So, you can see this prompt isn't very specific. It was pretty open-ended. I let it basically do its work, and I got out of its way. I obviously told it a few things like to
use my voice clone and my avatar and to, you know, work in some verification. So, you're not just giving me the first version. I think it probably would have done that on its own, but I always just
throw that in there at the end. Now, as far as the cost, I have been blowing through this limit like crazy. I've already used two of the resets that they gave us, and I'm 50% left on this third
one. But let me show you guys what this would have costed. If this was API billing, it would have been about 60 bucks. Now, I was also on fast mode, so it would have been more expensive. But
to get an output like this that took about 50 minutes is really insane. So, that was one of the first experiments I wanted to run. I'm obviously running a ton more. I've got a ton of other
content planned with Astra 6. I mean, I'm having so much fun here, but hopefully you guys enjoyed the video. If you did, please give it a like. Helps me out a ton. And as always, I appreciate
you guys making it to the end, and I'll see you in the next one.