Why I switched from Codex to Claude
I built Neighbourhood Finder with Codex. The backend, the web app, the data pipelines and supporting tools to process public data into neighbourhood reports. A couple hundred commits written in the evenings. Each commit or two being the result of a development session.
I use Claude all day at CybSafe, so I have a direct way to compare Claude and Codex.
I originally chose Codex because I could get the most done within the available limits. GPT-5 came out shortly after I started so the models were getting better. GPT-5.2 could write migrations that succeeded first time.
I wasn’t expecting to switch from Codex to Claude. Both Claude and Codex gave me what I needed. I was working on the assumption that Codex could give me more.
Then Anthropic released Opus 5.5 and OpenAI released Astra. I ask new models to review the active projects and methodologies. Astra gave some good insights based on the global view of success rather than at the individual project level. The insights felt deeper and were presented more clearly than before. Astra helped me work through a new idea, taking care of the desktop research and asking useful rubber ducking questions.
The problem was that Astra sucked through the session usage so quickly that I was constantly stuck. I’m usually doing 3 or 4 things at the same time in a session. At the same time, I saw an improvement in how much session usage Opus 5.5 consumed.
My plan was to use Codex and Claude together to expand my available usage and test what Opus could bring my projects.
The first real test was a short explainer video for that new idea. Before building anything, I wanted something I could show people to see if the idea made sense to them. I gave Astra and Opus the same concept and the same brief.
Astra was strong on the research. The video was different. It needed detailed feedback and many iterations to get close to what I wanted. Opus was better immediately. It used fewer tokens and the output was leagues ahead in polish and in the level it was working at. Opus went directly to what I was after and beyond.
Here are the two videos.
Fewer tokens matters more to me than it sounds. Fewer tokens is fewer iterations. Fewer iterations means I get more done in a session. More done is closer to my goals.
It’s hard to quantify the productivity increase I got with Opus over Astra. For anything creative, like video production, writing or website development, I’m getting about a 5x improvement. Most things Opus produces are good enough on the first attempt.
I don’t see any significant improvement in app development on either the frontend or backend. That is less of a loss than you might think. In a typical month, about 80% of my effort goes on research, content production and automating marketing tasks. Only about 20% is app development. The gains land where most of my time goes.
Very quickly I got bored with splitting work between Codex and Claude. I found myself using Claude for everything and cancelled my Codex plan. This step felt significant for me after using Codex for so long. It wasn’t the same as firing a teammate after hiring someone better, but close.