July 11, 2026/864 words/4 min read

GPT 5.6 Makes Codex Make Sense

The model got better, Codex got bigger.

OpenAI released GPT 5.6 this week and turned the Codex app into the new ChatGPT desktop app at roughly the same time. After using both, the timing makes sense. The model is much better at getting work done, and Codex has grown into the app that lets it do that work.

The naming is still confusing. The Codex app is now called ChatGPT. OpenAI added a regular Chat mode, though it feels pretty barebones right now, and a new Work mode that somehow seems even more incomplete. Codex is still there as the mode for software development. The previous ChatGPT desktop app is also still around, now called ChatGPT Classic, which more or less stamps "deprecated" directly on its forehead. Officially, OpenAI says it's not, and that it will keep getting model updates, security fixes and enterprise support. The new agent features are going into the other app, though, so we all know how this ends. Nothing gets Classic added to its name because the company has ambitious plans for its future.

The rename is a mess, but the new app makes sense because Codex has grown far beyond the tool I was using in April.

Back then I called it a pleasant surprise. It could work through a repository and run the checks without becoming a personality attached to a token furnace. I could give it a task, leave it alone and come back to review the diff. It was reliable enough to be useful and, most important, much less annoying than Claude had become.

Since then, Codex has gained access to much more than the repository. Appshots gave me a quick way to show it another application. The built-in browser goes further, letting Codex inspect the DOM, console and network traffic before using Computer Use to operate the interface.

Longer jobs are easier to manage too. Goal mode holds onto the task, while remote control lets me check in from the ChatGPT mobile app. One project can now span multiple repositories, and when the work is done I can review the pull request or edit the diff inline. That's quite a bit in three months!

Codex can now see and use more of the computer around the code. Before, it would change the implementation and wait for me to report what happened. I was the integration layer between the model and the screen. Every change sent me back to the page to reproduce the problem and carry whatever the browser told me back to the model. I even had to explain why a working layout still looked wrong. I could sometimes coax Codex into troubleshooting, but it was rarely successful. Now Codex can handle much more of that loop itself. It runs the application and tries the flow in the browser. If something breaks, it can inspect the console, return to the code and try again. It can also look at what it built instead of declaring victory because TypeScript stopped complaining.

GPT 5.6 is one of the best model releases in a while. OpenAI says it is better at coding and sticking with long-running work, especially when tools or Computer Use are involved. Fine, but that's what every model announcement says! "Look at this new model, best it's ever been!" Meh. I've learned the hard way not to trust that. This time I noticed the improvement before I read the announcement.

It stays on task much longer without hand-holding. I don't have to keep reminding it that the available tools are there to be used. It's also better at finishing the boring bits around the interesting problem. Earlier models would solve the main implementation and announce they were done while an error state or a test was still missing. GPT 5.6 is more likely to keep going until the task is complete.

Computer Use has improved too. Previous models could inspect a screenshot and recite what was in it. Heading, button, three cards, navigation bar. Great, thanks! They could see the screen without understanding why it felt unfinished. GPT 5.6 gets much closer. It notices when the spacing is off or an interaction technically works but still feels wrong, then connects that problem back to the code and makes a change. I spend much less time translating "this looks off" into exact CSS instructions.

Its judgment is better too. GPT 5.6 is less likely to turn a small change into an architectural decision or wander through unrelated files because it got distracted. It usually understands which part of the project I am asking it to touch and works with what is already there. I still review everything because it can absolutely invent details and misunderstand intent, but the review no longer feels like a search for a hidden disaster.

GPT 5.6 is probably the first model upgrade in a while where the improvement feels bigger in use than it did on the announcement page.

Codex is where that improvement becomes useful. GPT 5.6 can move between the repository and the running application, then keep working when the first attempt fails. I still review the diff, but I no longer have to operate the browser and report every result along the way.