GPT-6 Astra is a workhorse. Give it a directive and it keeps going: one four-prompt session building an infinite Backrooms game with a VHS look ran for about a day and a half, spawning subagents to check the work, add features and run tests.
That persistence is also its biggest weakness. Astra loves writing and running tests, even though the code it writes is strong without them. The game is live at backrooms.exe.xyz.
Summary
- Astra built an infinite Backrooms corridor game and a pixel shooter called Saffron Sands for me.
- It tends to write a test for each file it creates, then rewrite the test whenever the code changes, which wastes time.
- It handles harder work well, including tasks GPT-5.6 Sol refused or couldn’t finish.
- It reverse engineered the Bluetooth desktop app for my Amaran lights into a CLI and a custom skill, then changed and restored my lighting on command.
- It ran at about 76 tokens per second, excluding encrypted reasoning. It’s expensive, and it’s a better GPT-5.6 Sol rather than a giant leap.
How to apply it
- Use Astra for new projects and hard foundations, not everyday edits, because of the cost.
- Define outcomes first, such as the API, adapters and interfaces, and have the agent write one goal test it isn’t allowed to change.
- Let Astra iterate against that fixed test instead of creating and editing its own tests file by file.
- Point it at a clunky app or device you already own and ask for a simple CLI, then wrap the CLI in a skill so your agent can use it.
- Try Astra in OpenCode.
Verdict
Worth it for new projects and hard problems. Too expensive and too test-obsessed for everyday edits.
What’s good
- Keeps working through long, difficult tasks
- Handles work GPT-5.6 Sol refused or couldn’t finish
- Reverse engineered a Bluetooth app into a CLI
What’s not
- Expensive for day-to-day use
- Writes and rewrites tests compulsively
- An upgrade over Sol, not a leap
Get the next test the day it’s out.
