Over the last week I've been experimenting with OpenAI's Codex app, building a family calendar application. It's been a great way to learn more about AI software development, and I've ended up with a useful app for our family.
The app runs on GCP and replaces a shared document we've been using for years (though it can still export to PDF in the old format). It's multilingual, automatically imports calendar feeds from several different sources such as the soccer teams and Bulgarian school, and does pretty much everything I wanted it to do.
Codex with GPT-6.1 Sol
I mostly used GPT-6.1 Sol Medium, which has been terrific. I also tried Astra in the initial phases, and it was excellent, but burned through too many tokens for my $20/month ChatGPT Plus subscription. By its own estimate, I used in the ballpark of 10 million tokens for this app; GCP running cost would probably be $3 per month at the most.
Codex worked incredibly independent. It routinely spent 10–15 minutes coding, testing, and fixing things without any input from me. That's a very different from working with AI in VS Code just six months ago, when I still had to do a lot more hand-holding.
The funny thing is, I barely know the architecture of the app. I know it's a 1-vCPU OCI container running Debian in Google Cloud Run, but Codex set up all the GitHub build pipelines and deployments for me. I did have Codex spent quite some time on security scanning and building, and having automatically pick up patches etc.
It's just amazing to see how quickly things are developing.
Comment