ChatGPT 5.5 Codex: I can't believe they did this...
TL;DR
Alex Finn says Codex desktop is the real story, not just GPT-5.5 — he claims the new app is where he now spends “90-95%” of his day because it combines coding, chat, browser, annotation, and computer control in one place.
GPT-5.5’s biggest upgrade is that it actually takes action — compared with GPT-5.4, Finn says 5.5 is less verbose, more pleasant to use, and smart enough that he’s seen “zero bugs” in the code it wrote during his recent testing.
His recommended default setup is surprisingly simple: medium intelligence, fast speed, full access — after testing low through extra high, he says medium is the sweet spot in Codex, unlike Claude Code where he typically cranks everything to max.
The killer workflow is image-first app building — using OpenAI’s built-in image generation inside Codex, he has the model generate 4-5 UI concepts first, then build from the chosen mockup, something he says Claude Code can’t match because it lacks integrated image gen.
Computer Use and live annotation are what make Codex feel like an AI employee — Finn shows Codex editing UI elements directly in the browser and then autonomously testing the app by moving the mouse, clicking features, and reporting its progress while he multitasks.
His advanced stack is Codex + Linear, and his verdict is a full switch away from Claude Code — by letting Codex create and manage Linear issues across multiple agents, he says he can build “10 times better and faster,” and concludes GPT-5.5 Codex has finally dethroned Opus 4.7 and Claude Code.
The Breakdown
Why GPT-5.5 immediately felt different
Finn opens hot: he’s been using ChatGPT 5.5 nonstop for days and says he’s genuinely blown away. The headline change for him is simple but huge — 5.5 “takes action by default,” unlike 5.4, which he jokes made him feel like he was begging on his knees just to do the thing.
Codex desktop becomes the main character
He says the desktop Codex app is the best way to use GPT-5.5 and urges viewers to pause the video and download it before doing anything else. What wins him over is the tight integration: projects on the left, coding and planning chats in the same app, and less of the awkward mode-switching he dislikes in Claude’s separate chat/code interfaces.
The setup he actually recommends
Before building, he points to two settings that matter: turn on Computer Use and stop overthinking intelligence settings. After testing low, medium, high, and extra high, his conclusion is that medium is the sweet spot for most work, and he prefers full access permissions so he doesn’t have to keep clicking approve every few seconds.
Building a prompt library app, starting with image generation
His demo app is a prompt library that saves, favorites, and organizes reusable prompts, but the key move is starting with UI exploration instead of code. He has Codex use OpenAI’s image model to generate five interface options, picks the green-accented card layout, and frames this as a major advantage over Claude Code, which he says simply cannot do this because it has no image gen built in.
Annotation and browser tools make the app feel alive
Once the app is generated, Finn highlights Codex’s built-in browser and annotation flow. Instead of writing another long prompt, he clicks directly on the interface, marks a location, and asks it to add a favorites button — and the change appears right away, which he says feels more like building inside the app than chatting with a bot.
Computer Use turns testing into delegation
Then he shows off the feature that most excites him: telling Codex to test the app itself. The browser opens, the AI-controlled mouse starts clicking, typing, saving, and favoriting without his help, and he describes it as having an “AI employee” running QA while he goes to eat lunch or works on something else.
His advanced workflow: Codex plus Linear as a second brain
The next level, in his view, is pairing Codex with Linear’s free tier. He has Codex connect to Linear, create a project, generate around 10 issues with categories and details, and then use those shared tasks across multiple sessions so different agents can work side by side without needing to be re-briefed every time.
The Claude Code verdict finally flips
Finn is explicit that he has been a “Claude Code guy” since March, but says this is the first time in over a year that the crown changes hands. His case for GPT-5.5 Codex is better app experience, higher limits, more consistent intelligence, integrated image generation, and Anthropic’s apparent compute constraints; he still gives Claude credit for better default UI, warmer personality, and slightly stronger stubbornness, but his final verdict is clear: for now, ChatGPT 5.5 Codex is the best coding stack.