ChatGPT 5.5 Codex vs Opus 4.7: The winner is clear
TL;DR
Alex Finn says ChatGPT 5.5 + Codex has overtaken Claude Code for coding — after “hundreds of tests” against Opus 4.7 and Claude Code, he says the combination of the smarter model, better app experience, and much higher limits makes the winner “clear.”
The Codex desktop app is the real story, not just the model upgrade — he says he now spends 90-95% of his day in the app because it combines projects, coding, separate planning chats, a built-in browser, live annotation, and computer control in one place.
His practical setup is surprisingly simple: use medium intelligence and full access — unlike Claude Code, where he tends to max everything out, he says ChatGPT 5.5 is most effective on medium, and turning on “computer use” plus full permissions makes the workflow dramatically faster.
The biggest differentiator is integrated image generation and UI prototyping — he uses OpenAI’s new image model to generate 4-5 interface concepts for a prompt-library app before building, calling it “by far and away” the best image model and something Claude Code simply can’t match.
His advanced workflow is Codex + Linear as an AI project manager stack — by connecting Linear, Codex creates projects, adds detailed issues, and lets multiple agents work off the same task board, which he says made him build apps “10 times better and faster.”
Claude still wins in a few narrow areas, especially default UI taste and personality — Finn says Opus 4.7 still feels warmer, more concise, and better at producing polished default interfaces, but he thinks Anthropic’s compute limits are causing the model to get “significantly stupider” over time.
The Breakdown
Why 5.5 immediately impressed him
Alex opens hot: he’s been using ChatGPT 5.5 nonstop for a few days and says it feels fundamentally different from 5.4. His biggest praise is that it “takes action by default” instead of needing to be begged into doing the actual work, while also being less chatty, a bit warmer, and much smarter at writing code.
Codex desktop became his default workspace
He’s adamant that the desktop Codex app is the best way to use 5.5, to the point that he says he now spends 90-95% of his waking time inside it. The pitch is integration: projects on the left, multiple sessions inside each project, and separate chat threads for planning and strategy, instead of Claude’s split experience where coding and chatting live in different interfaces.
The settings that matter: medium intelligence, full access, computer use
Before building anything, he tells viewers to flip on “computer use” in settings because it lets Codex control the whole computer and test apps on its own. His other strong recommendation is counterintuitive: don’t run extra high intelligence all the time — after testing low, medium, high, and extra high, he says medium is the sweet spot, paired with full access so you aren’t constantly clicking permissions.
Building a prompt library app starts with image-generated UI concepts
For the demo, he creates a prompt library app and leans hard on the built-in image model to generate five possible interface directions first. That’s one of his key knocks on Claude Code: it doesn’t have integrated image generation, while Codex can go straight from “make it simple, beautiful, like Linear” to visual options, then immediately build the selected green-accent card design.
The killer UX: annotate the app live and let AI test it itself
Once the app is running in Codex’s built-in browser, he shows his favorite trick: clicking directly on UI elements with the annotate tool and asking for changes like “add a favorites button” instead of writing another full prompt. Then he hands testing over to computer use, which opens the app, moves the mouse by itself, clicks around, fills forms, favorites prompts, and effectively acts like “your own AI employee” while you go eat lunch.
The advanced workflow is Codex plus Linear as a shared second brain
His “life-changing” workflow is pairing Codex with Linear, which he calls “peanut butter and jelly.” Codex can create a project in Linear, generate issues with details and categories, and then multiple agents in separate Codex sessions can all work off the same board, eliminating the constant ritual of re-explaining the codebase and next steps.
Where Claude still wins — and why he thinks it’s slipping
Alex doesn’t pretend Claude lost every category: he still thinks Opus 4.7 has better default UI taste, a warmer personality, and is a touch more proactive. But he argues Anthropic keeps losing ground because its models feel worse after launch, limits keep dropping, and the company appears compute-constrained, while OpenAI’s bigger infrastructure bet means 5.5 stays consistently smart and usable.
The verdict: after a year, there’s a new coding champion
He frames this as a genuine changing of the guard, since he’s been “a Claude Code guy since day one” and had basically stopped using everything else. But at this moment, he says the combo of ChatGPT 5.5, the Codex app, integrated image generation, computer use, and higher limits makes it the new best way to build apps — at least until the next model shift forces him to switch again.