Matt Shumer published a hands-on review of GPT-6 Astra, OpenAI's new model. He calls it his default for almost all work, ahead of Anthropic's Fable 5. Backend engineering and computer use are the strongest areas, he writes. He says Astra replies in plain English, which makes running several agents easier. Medium reasoning effort suits daily work; Ultra suits larger builds. Claude still wins on visual taste and asset creation, he says. He describes a setup he calls the Manager Loop. One agent coordinates while a second implements in a separate Codex session. He used it for a simulated civilization and an Unreal build of New York. He says long-running autonomy is not solved and runs still plateau.
What changed
Earlier OpenAI models needed close supervision; one deleted nearly every file on his computer.
What it unlocks
Running one long build from a single prompt, with a coordinator agent driving a separate implementer session.
- up to 96 sub-agents at once
- default Codex limit of 4 sub-agents
- 16 sub-agents on a second machine
What you need to act on it
- access to GPT-6 Astra
- Codex
- changed Codex sub-agent limit
- large token budget
Sources