Peter Gostev OpenAI Tests GPT-6 Astra

Peter Gostev's OpenAI test of GPT-6 Astra shows a model that can migrate 150,000 lines of vibe-coded legacy code and render a voxel map of historic London without constant hand-holding. There's a catch, though. It only works when run remotely on a Linux box, since a local CPU becomes the bottleneck.

StartupHub data

Companies working on this

Profiles of the companies named in this story, with funding and a one-liner from our database.

OpenAI

$122.0B

Artificial intelligence research and deployment company focused on developing advanced AI models like GPT-5.6 and GPT-Live, offering products such as ChatGPT, DALL-E, and Whisper.

In a video on OpenAI's YouTube channel, Gostev explained that he built the London map to push 3D visual capabilities across medieval and Tudor eras, then threw his hardest tests at Astra to see what would break.

The full discussion can be found on OpenAI Youtube's YouTube channel.

How the new behavior actually works

The previous system, Sol, would instantly agree with any criticism and tell the user they were right.

Astra pushes back naturally, admits when it's wrong, and keeps exploring the solution space longer on difficult tasks.

Think of it as the difference between an intern who asks for approval on every line and one who rewires the whole house correctly while you're out.

What works, and what still hasn't been fixed

For builders, the appeal is reliability on large, messy codebases that piled up since the GPT-5.2 era, code that previously needed extensive debugging after GPT-5.6 runs.

The bottleneck now isn't model quality but compute placement. Gostev had to move the work off his laptop to a remote, plugged-in Linux box so his battery and CPU wouldn't die mid-task.

What nobody's answered yet is how Astra handles permissions, sandboxing, or recovery when that long-horizon autonomy goes sideways on production systems.

The model has grown up. Your laptop hasn't.