Update on the project comparison (created a "timekeeping" app attempted in a simultaneous one-shot session with Claude Code Opus4.6 and GPT5.4): I hope this is helpful. I think immediately a clear winner emerged. Give feedback and let me know if I left something out. I gave the exact instructions with identical plans (created by Grok for fairness) to both models. Here's what happened: -Both finished the first pass in about the same time, within a few minutes about 40 min total. GPT produced in stages so a little more time 20 min was required notes below. -GPT produced a "initial MVP" purely for viewing and did this several times. Rather than delivering every step at once, GPT allowed confirmation at several steps. I like this approach. Catch errors in stages. Took extra time, but it was welcomed. -The first pass was more complete from Claude, but not everything worked refactoring required on most of the windows, especially errors involving CRUD operations. I had to request refactoring several times, still trying to resolve CRUD issues. Not resolved. -I actually liked the frontend GPT5.4 generated, it's actually very pleasing to use. -Claude had several parts that did not include add, edit or delete options. I could not add team or clients. After the "Add" buttons were included, they did not work. -Tried to enter a new client in Claude, it would not add. It froze, then the record did not add, same for projects. -GPT5.4 is building everything in steps and requests checks as we go along, I kind of like this feature. So far everything is working but the next phase is the CRUD, where problems are usually found. Essentially no errors. Summary with winners (sorry not in a table): Frontend design: GPT Fully functioning: GPT (after stepping through phased output), everything worked no issues CRUD: GPT, Claude is still not ready, taking several passes and still does not work. I'm not going to keep trying. Fields included: GPT was much more thorough with all the fields as specified in the design plan. Claude took the minimal approach, left fields out. Followed the design: GPT, I'm still checking this, but appears every design element was followed. Claude omitted many items. Code review: Tie - this was done at a high level but the code seemed cogent without strange patterns, linting was self checked by both, and no problems were found (not the best analysis, but fine for purposes here). Token use: Claude 45k, GPT recorded zero tokens, this has to be configured up front and I didn't do this. Apologies on this, as mentioned I have the max subscriptions on everything so I usually don't pay attention to the costs, unless I hit a wall, which we rarely do. Winner analysis Here's my assessment. GPT5.4 clearly won this one. It followed the design, gave steps for checking and after adding CRUD operations everything worked without error. Plus the frontend design is so pleasing and easy to work with in GPT. I would actually enjoy working in the GPT app and I might end up using this internally. Not sure about making it a product offering though. I'm a huge general Claude fan, I use it for many things, but Claude Code (CLI) needs work. I think Opus 4.6 is smart enough, but for some reason it does not produce rock-solid out of the gate. This has happened with several projects. Screen shots: Claude's first pass (the plain looking one), and GPT's first pass (the fancier looking window); followed by GPTs project entry window followed by Claude project entry window. I'm not sure how the images will show, but GPTs windows have much more verbiage and has the very light tan background. Claude images are black and white.

I'm testing Grok Build with a new project. I'm not a huge fan of CLI, but it's powerful and in this case Grok is doing a great job so far. Super fast at this point. I'll report more as progress is made.

So a very large customer says "hey we cannot find this functionality out there, can you build it???". Of course I said "yes". How could I say "no" to my new project addiction. But this one already has a need!

Building a nice specialty software for scheduling/timekeeping for first responders using GPT5.5 (med). The model has performed perfectly. I've asked the team to try to estimate the time savings and they estimate about 75% to 80% time savings for coding, maybe 25% unit testing time savings. No real savings in code review and UI design. The last two parts cannot be omitted.

Already up and Codex (GPT5.5 medium) is running through "slices" 99% of the time without error. This is great. It still needs SWE interaction and testing/code checks, but I think over all it saves about 60% in coding time. That's significant in terms of delivery time, NOT employee replacement time.

Had a customer tell me they decided to create their entire accounting system and warehousing system using AI. I told them best of luck, it's not as easy as it might seem especially with very complex requirements. We will see if they can pull it off. We will be here to catch them if they fall. 😂 But we will be observing.

Codex (GPT5.5) seems to be improving with its UI skill. I'm actually impressed by some of the screen layouts provided. previously, the model would ignore our "screen layout rules" and use its own structure. It now follows every rule and consistently produces UI as planned. Seems the improvements are being updated and improved without a lot of hype and fanfare. That's good.

Working with GPT5.5 this morning and I'm impressed with the UI creation. Seems the UI skill is being improved. The layout is actually very good and is time saving. Sometimes it still adds unnecessary UI documentation like "Click this button---it no longer produces an error like before." 😂😂

I get lots of email advertisements that are obviously AI created. Should I care? Do I hold them in contempt because they used AI to create their spam email? It's fine. I don't care. Maybe my AI email is deleting their AI spam...

One of the things I require of my SWEs, is to apply all migrations manually. This way, they can review them before running, understand them, and make revisions where necessary. This comes in very handy when you are applying migrations to a few hundred production database. Whew.

Working with GP5.5 on a new project and noted that the UI skill has improved significantly. This is great. We still get a UI expert to review, but this saves lots of time.

companies might reduce their workforce because of AI, but they likely a) already had too many employees and they are RIFing because it is needed b) don't understand that AI is a complex beast, accountable to no one, and they will regret the layoffs.

We are creating another app (we can't help ourselves, it's an addiction). Customers are requesting it, so it has validity. I gave the idea to 4 different AI's for analysis and ideas: O4.7, G5.5, Gemini and Grok. I will say GPT5.5 gave the most comprehensive analysis, Followed very closely by Claude, Grok then Gemini. I continue to be impressed by GPT5.5!

Product shipped! We had help from Codex on this one. It's a specialty payroll integration product so it's not for every business. We converted a long-time desktop product to a SaaS app and customers are eager to switch to it. Overall, I think we probably cut our dev time by about 50% using Codex. We still had to review all the code, make a bunch of SQL scripting changes for the migrations, test upgrades on a bunch of tenants. But it works very well. And the UI is being well received by our end-users. Codex has earned that employee of the year parking space.

Coding and shipping today! Codex just helped get a new payroll product out the door. I tried using Claude this morning to do some planning and it would not even upload a small document that was to supplement the planning process. However, Codex accepted several very large files and provided a super good plan. I'm trying as hard as I can to keep using Claude, but it's making it very difficult.

I'm finding Opus 4.7 to be degrading (the coding side). It's like a repeat of O4.6, is anyone else seeing this? For the first week or so, it produced stellar results, now it skips details in plans and design specs, inserts shortcuts that don't really make sense. The planning and document generation is still great, even with really detailed charts and graphs, but the coding is like a different model. It takes 5 or 6 iterations to get something right, when previously it would nail it on the first try.

I've been coding since early this morning, I've really giving Opus4.7 an honest try, it does understand some of the UI components better, but it skips prompting instructions way too much, which requires sometimes 9 or 10 iterations to get it right, even when I provide explicit instructions to follow a template that we designed together. The planning capabilities are great, but the coding seems to be a different model that just tries to finish too quickly without all the required pieces. I always add, "please follow the design template strictly", Opus will do part of it, but then stop, or create a totally different process that is not consistent with the other UI components. I'm not sure what causes this, maybe the context is too big, (the UI is has been repeated throughout) ? I'm sharing this to be constructive, not to badmouth Opus4.7. I think we need very competitive tools throughout. Codex (G5.5 high) is not flawless, but the end result is 1 iteration, maybe 2 and things are correct (usually the requirement of the second iteration is my fault for not being clear enough and maybe I misread the plan that Codex provides before the code changes). Seems like these guys have provided a much better tool.

Building a product and I had to do some SQL scripting. It feels good to actually write some SQL Scripts for a change, rather than letting AI do it. I feel such nostalgia!

Currently I'm using Codex (G5.5 medium) to build a semi-complicated integration for an ERP system. I suggested a design change GPT came back with a much more extensive, forward-thinking plan. I like that. It did not "go ahead and create" the new ideas, it paused and proposed the new plan, with very good future addon capabilities for me to approve. This is proper procedure for AI as a "team member".

Up and coding early. We are shipping product using GPT5.5. It's a workhorse and very dependable. Let's hope it stays like this!

I gotta say, GPT5.5 (medium) does a significantly better job at creating UIs! I'm duly impressed. Nice work OpenAI team! Obviously, humans do a better job at UI, because there a nuances that aren't readily apparent to an AI model, but this a huge improvement!