I Gave GPT-6 Astra One Prompt. It Fixed Its Own Mistakes.

0:00 12:26 in the afternoon. I paste one brief into Higsfield Supercomputer with GPT6 Astra picked as the brain and I take my hands off the keyboard. Watch the clock in the corner because the

0:12 clock is the whole test. Every review of GPT6 Astra so far is a conversation. Somebody types it answers. They grade the answer. That's the wrong test for a model whose whole pitch is doing things

0:25 on a computer. So, I gave it a job and I don't even have chat GPT+. Here's how this goes. First, what Astra is in numbers. Then, you watch it run a real ad campaign on my Higsfield account from

0:38 one prompt. Then, it builds a 3D set. And at the end, we read its bill against my real balance because that's where an agent either lies or doesn't. GPT 6 Astra shipped on the 3rd of September,

0:50 rolling out to the paid tiers over the following days. Astra Pro is a separate tier and it isn't on plus at all. The context window is about a million tokens and Higsfield is one of the named launch

1:02 partners. The headline feature is computer use and Open AI means it. The launch demos have Astra laying out a circuit board and modeling a house in Blender. On the desktop benchmark, it

1:14 beats the previous model and it does each task in about half the time. Now the skeptics receipt because you'd ask. The 99.9% ARC AGI three score in every headline

1:25 came from OpenAI's own custom harness. And that run cost $19,000. On the ARC prizes standard harness, it's 62.7. Brilliant at thinking. Thinking isn't making. So that's the model on paper.

1:39 Now the tools it gets. Inside Supercomputer, the model picker list auto, which is free. Then GPT6 Astra tagged high cost about 17 credits for a typical message. You pick Astra. You

1:52 pick an effort level. Then the switch that matters. Generate without asking or ask before every generation. I set it to run without asking. Think of it as the coding agent loop you already run in a

2:04 terminal except the tools are cameras and models. And one caveat, Higsfield states plainly, "Agent generations charge credits on every plan, unlimited included." That's the setup. Now, the

2:16 job. Here's the brief. The only thing I typed in this run, a 10-second hero clip for a product ad, plus two poster stills. Work like a creative director. Read my balance first. Plan the shot in

2:29 three lines. Pick the models yourself and tell me why. When the clip is done, look at it. Tell me what's weak. Rewrite your own prompt and go again twice. Then bill me. I had no unbranded product to

2:42 photograph. So the brief opens by inventing one. A cold brew called Kestrel. A matte black can. No real logos anywhere. 19 seconds in the first tool call. Astra reads my balance back

2:54 to me. 2,200 credits. Remember that number. Then the plan. Three lines exactly as asked. Subject: Kestrel Cold Brew. An invented matte black can with ivory lettering and one copper slash.

3:09 Camera, a slow, low three quarter push that settles into a readable hero frame. Light, a narrow, cool rim and a warm grazing key. I typed none of that. Then the models its choice Nano Banana Pro

3:23 for the stills for crisp packaging typography and seedance 2.5 for the clip for controlled camera motion and consistent product geometry across three takes. Then it wrote the video prompt

3:34 itself. A camera 85 cm from the can, a 29° field of view, a 15 cm push, a single bead of condensation creeping down under gravity, a 2- second hold. The best prompt ever run on this account

3:49 is one I didn't write. 12:30. The pack shot lands in the assets panel on the right. Then the two posters. And here's the first catch. The bright poster came back with an extra word in the headline.

4:01 The word exactly sitting right before take flight. Astra spotted it before I did. It's exact words. The bright poster picked up an unwanted word in its headline. I'm correcting the typography

4:13 without another image generation, which means it fixed the file instead of paying for a roll. Two credits saved. Sure. The reflex is the point. 1232 version one starts rendering. 10 seconds

4:26 of cedance at 1080p and Astra doesn't fill the silence. Its note says, "I'm waiting for the actual footage before deciding what to change. I expected it to pre-write the critique. It didn't."

4:38 Version one comes back and Astra looks at it with a video analysis tool, not a guess, verbatim. The can's lower wall pinches and its base rim changes shape midshot, undermining the rigid metal

4:50 feel, and the push drifts upward and sideways. The move feels floaty rather than mechanically precise. Watch the base, so it rewrote its own prompt. Version two, keep the can rigid and

5:02 stationary. One straight fixed height dolly push. Lock the label's position. finished with a two- second hold. Then it started version two. 1240. Version two is in. And Astra isn't happy with

5:15 that one either. The highlight keeps sliding across the matte face and the camera doesn't settle before the ending cuts the push off. Version three, softer camera relative lighting. Complete the

5:26 push by six seconds. Decelerate by eight. Hold still through 10. Here are all three side by side. Version one drifts. Version two slides. version three holds, not one note typed between

5:39 them. So, who's the director here? That's the recursive loop from the launch page, and it only works because the model can see what it made. While version three renders, it does something

5:48 I didn't ask for. It writes itself a supercomputer skill called three take product direction. So, the next product ad on this account starts with the method saved. Then it hears an unrelated

6:00 pouring sound on the clip, strips it, and exports all three takes at exactly 10 seconds. Now the bill 12:46, 20 minutes after I pressed enter, the thread renames itself Matte Black Cold

6:13 Brew, and the bill arrives. Packshot, two credits listed. Each poster two, and each hero take a balance drop of about 96 credits, the exact figures on screen. And then the line I respect most. The

6:27 video figures include chat and tool usage. They are not isolated generation fees. Starting balance 2,200.18. Total observed decrease 389.88 and an exact generation only total could

6:42 not be verified. It refused to guess. So I checked. My account after the run reads 1,85.88. That's about 4 1/2 credits more than Astra's figure. Most likely the export

6:54 in the last tool calls after it last looked off by four credits and it had told me why it might be. That's the ad. Now the second job, a different tool in 3D Judsu. You prompt a scene, get an

7:08 editable 3D scene, then transform it into video. The model chip list GPT6 Astra 2. So I picked it. One prompt. A small product studio. A low pedestal. One softbox. A curved white backdrop. A

7:23 matte black can. Camera slowly orbiting. A couple of minutes later, it's a real scene, not a picture of one. A camera object, a softbox, a backdrop, the can on its pedestal, all in a properties

7:36 tree you can click through, and you can orbit the whole thing. Then you render and the page opens Cinema Studio 4 by itself with the render attached as a reference. I typed four words make this

7:48 add cinematic and 45 credits later listed at 80 on the button there's a 5-second cinematic of the can and smoke. The can by the way is still Kestrel. If you live in chat GPT instead the same

8:01 hands plug in through Higsfield MCP the server behind all of this on my free account. The plugin took a Higsfield signin and a brief and came back with a style frame and four questions. Astra

8:14 there needs plus which I don't have. So that brain was the default model. Now the part I didn't script. This was my first time inside supercomput. For months I've made assets for this channel

8:25 one at a time each with its own prompt. About half an hour of typing for a set like this. Here I gave direction what to do and what not to do and it wrote the rest. If you build things and then have

8:37 to market them, that's a different kind of leverage. Here's where I land. GPT6 Astra didn't get smarter inside Higsfield. It got hands. And hands change what you can judge. Judge an

8:48 agent by what lands in your assets folder, and by whether its bill matches yours on this run, six finished files, two self-critiques, one saved skill, one typo caught, and a bill within five

9:00 credits of the truth. The limits are real, too. Astra is the high cost option in that menu. Agent generations charge credits on every plan, and a 10-second take moved my balance, about 96 of them,

9:13 chat included, which leaves the thing I keep chewing on. If the best prompt on your account was written by the model, and the best critique of its work came from the model, what exactly is your job

9:24 in the loop? That's the take. Higsfield sponsored this one and the link to try supercomputer with Astra or the Higsfield plugin in chat GPT is in the description below. Hand it one brief and

9:36 read the bill.