Codex has now passed 𝟭𝟱 𝗺𝗶𝗹𝗹𝗶𝗼𝗻 users and grown 𝟮.𝟱𝘅 over the past month. That pace is clearly taking market share from Claude Code. I still use both Claude Code and Codex heavily. On model quality alone, Opus and Sol are very hard to separate. Overall, I’d call it roughly even. Day to day, though, I’m almost entirely in Codex. I mostly use Claude inside Herdr fleets, where it joins the other agents for something closer to an internal team discussion. 𝟭. 𝗖𝗼𝗱𝗲𝘅 𝗶𝘀 𝘁𝗵𝗲 𝗮𝗽𝗽 𝗜 𝗼𝗽𝗲𝗻 𝗲𝘃𝗲𝗿𝘆 𝗱𝗮𝘆 The Codex App is just smooth. I can queue prompts, sessions are easy to find, and switching between them never gets in the way. Computer use and browser use are much more reliable. Claude still fails midway through a run pretty often. Codex even notices whether the cursor is visible enough to use. For browser work, I recently started preferring ego-browser (ego-lite). Worth trying. I worried about Codex’s smaller context window. After multiple compactions, it still remembers what we were doing. Claude sometimes compacts and feels like an entirely different person. I also have Codex clean up my agent skills from time to time. First it inventories them without touching any files: every skill goes under KEEP or REMOVE, with a one-sentence reason. I read the list, then Codex deletes only the REMOVE items I approve. Duplicates, stale skills, and irrelevant ones surface fast. /goal is useful too. The session knows when to pause, and the goal stays visible in the App. I don’t have to explain it again. 𝟮. 𝗪𝗵𝗲𝗿𝗲 𝗖𝗹𝗮𝘂𝗱𝗲 𝘀𝘁𝗶𝗹𝗹 𝘄𝗶𝗻𝘀 Fable is still very wise. Claude’s hooks go deeper. When I want fine-grained control over how a session behaves, there are still plenty of details Codex can’t touch. Routines and Managed Agents are more mature too. The Team plan also offers Premium Seats, which makes Claude easier to run across a team. 𝟯. 𝗧𝗵𝗿𝗲𝗲 𝗺𝗼𝗿𝗲 𝗿𝗲𝗮𝘀𝗼𝗻𝘀 𝗜 𝘂𝘀𝗲 𝗖𝗼𝗱𝗲𝘅 Image generation saves me a ridiculous amount of time. Claude has barely touched this. Gemini’s Nano Banana Pro used to be my favorite, and for a while I thought it was the best. Codex image generation replaced it in my workflow. Nano Banana Pro’s output isn’t as good as GPT Image 2, and Gemini is still terrible at coding. I’ve also been testing subagent setups a lot lately. I made Luna Max my default subagent, then removed it a few days later. It’s fast, but open-ended development creates enough rework to erase the time it saves. I now default to Sol Medium. Hard tasks go to Sol High. For small tasks with clear boundaries and easy checks, I still use Luna. Haiku often isn’t usable on the same work. I tested several models on one very specific step. Before image generation, a system prompt has to interpret the user’s instruction and write the processing prompt that comes next. Luna kept beating Sonnet 5 and Gemini Flash 3.7. It was faster and more reliable across repeated tests. Sonnet 5 costs more. Weird result, but I kept getting it. Codex App Server also fits how I work. The spec is open, it’s easy to hack on, and there aren’t many restrictions. If I could keep only one today, I’d keep Codex. For real projects, I still keep both open. Claude and Codex get one pane each. One builds, one reviews, and the main agent sits in the middle assigning work, waiting for results, and following up. I used to handle every handoff myself.