Claude feels mode-first. Codex feels work-first. That distinction mattered more than I expected.
I used Codex with GPT-5.5 for about three weeks. A colleague asked me for notes on my experience with Codex at the end of this time frame. When I started to write up my thoughts on the experience, I discovered that it was not the model itself but rather the overall experience of the application that was most important to me.
Claude features different modes for different kinds of tasks. Depending on the nature of the work I want to do with Claude, I can choose a mode to work within. Codex, however, does not feature such a mode selection system. Instead, I find myself jumping straight into Codex with no decision required on my part about which mode to use to perform my desired task. The application automatically jumps me into a mode that can perform the work I need done or allows me to initiate the various stages of that project directly within the application. For example, I no longer have to open up ChatGPT separately to answer quick questions.
My typical day as an FDE involves working with customers, working on bug bashes, writing documentation, and coding. Each of these tasks happen in sequence and in varying frequencies throughout the day. While using Claude I typically have to make a decision of which mode to use to complete each of these types of tasks. While individually they may seem unimportant, the fact that I have to make these decisions dozens of times a day is an element of my workflow that differs from Codex.
The built-in browser in Codex was especially helpful for my type of work. During bug bashes, for example, I can highlight portions of a website and include those elements in the task that I am working on in Codex. I would have otherwise had to take a screenshot of the portion of the site I was highlighting, paste it into Codex, and tell Codex what portion of the site I was looking at. Now that functionality is built into the product.
GPT-5.5 was a solid chat model. The best way to use it is by prompting desired outcomes and success criteria. It maps better to how I think than some of my recent Claude prompting.
Overall, for the kind of work that I do, the fact that Codex requires no decision on my part about which tool to use for any given task fits my workflow the best. Claude is better for when I know in advance the tasks that I will have to perform.
While neither model is obviously better than the other, it is clear that the experience of using Codex versus Claude has a bigger impact on me than the models themselves. At Airbyte, we think a lot about how agents interact with data. It’s clear that the surface around the agent has an equally important impact on how people interact with the system.
This is the fifth post in Deployed with Dallas, a series on building AI into my day-to-day as a forward deployed engineer at Airbyte. Expect more posts soon.
