Quick confession. Last Sunday I opened dev.to and my brain made a tiny error noise. 😵
Opus against GPT-6 against Gemini 4. One post said Gemini wins twelve of eighteen benchmarks. Another said a Meta model costs the same as Opus and loses most of its own tests. One minute of scrolling. Zero joy.
So I made a deal with myself. Thirty days. One editor, VS Code. One model, Claude Sonnet 5.5, on the free plan. Yes, free. 😂 Every new launch was allowed to happen without me.
My old routine, for context. Google AI Studio on Monday, trying one Gemini flavour after another. DeepSeek on Tuesday because a bug felt like its kind of bug. Then I'd test a few models inside my AI powered projects too. I called it research. My calendar called it wandering. 🧭
That same week, FrancisTRdev's ( @francistrdev ) post about AI fatigue was sitting in my feed. He misses the dev.to from before 2022, the one with CSS tips and small weird projects, and he hopes AI becomes a quiet boring tool in the toolbox. Boring! That's the exact word I wanted for my editor. His post got well over a hundred reactions, so I'm clearly not the only tired one.
Sylwia Laskowska ( @sylwia-lask ) wrote about AI features that run in the browser and what your hardware can really handle. Jamilxt ( @jamilxt ) keeps pointing out that a model priced at two dollars per million tokens can still give you a very different real bill. 💸 Maame Afua A. P. Fordjour ( @maame-codes ) wrote that localhost is a lie, because real users break clean databases. Three different voices, one message. Real work beats shiny news.
So what did my thirty days look like?
Quiet. Busy, but quiet. I finished a few client projects, some for clients in India, many of them ecommerce sites. I built the website for my agency TheBitForge ( @thebitforge ) and worked on its projects. I wrote a few articles for topblogs.online and some other places. And after a few emails with the founder of Floot.com, I used a paid Floot plan connected with Claude to build a full stack blogging platform with an admin panel and AI summaries for every article, plus a newsletter system for my own use.
Claude Sonnet 5.5 helped me well through all of it, the ecommerce work included. And look at that list again. Where's the new model? Nowhere. I just finished things.
Now the part that hurt.
One client got their ecommerce website from me. They didn't like it. A few days later they came back and asked me to redesign the whole thing again. All of it. 😅
That was easily the worst moment of my month. And here's the strange part. No model on earth could have stopped it. Not the newest one, not the priciest one, not the one with twelve benchmark wins. A client's taste isn't inside any benchmark. Maame was right. The hard part of this job is people, not tokens.
What I'm taking from it? Show the client a rough version sooner, before the pretty one. Ask what they hate before I build anything. Simple, a bit embarrassing that I needed a full redesign to learn it. 🙃
Do I think everyone should stay on one model? No. If your job is literally comparing models, please keep comparing. For the rest of us, I'd check the news once a month instead of once an hour. If something is truly better, it'll still be better in four weeks, and by then someone on dev.to has probably tested it and posted the real bill.
Your turn. Has a client ever asked for a full redesign right after you delivered? How did you handle it? And which model are you on right now? Tell me below, I'm genuinely curious. 👇
Top comments (6)
Thanks for the shoutout, Hanzla! Really glad the localhost post stuck with you.
The model fatigue part is so real. Sticking with one and just building things sounds like a nice reset. Sorry about the redesign though, but “show the rough version first” is a lesson I’m keeping too.
Haha yeah, I learned that one the hard way 😅 Glad it stuck with you, Maame!
:)
The bill point is the one I'd check first. One tester reported about 1 percent of their plan per task on Sonnet at medium effort versus about 20 percent on GPT 6 Astra at high effort. Same kind of task, very different usage.
The redesign request is the part of this post I suspect most people here recognise. No model would have prevented it, as you say, because the gap was never in the code. It was in what the client pictured before the first line was written.
Your two takeaways, show a rough version sooner and ask what they hate before building, are good ones. I would add a third: write down what the site has to do for the client's customers, in plain words, and get a yes on that page before any design starts. For example, a returning buyer finds last month's order in two clicks, checkout works on a phone, and the owner can add a product without calling you.
It gives both of you something to point to when a full redesign request arrives, and it makes it easier to tell a change of taste from a missed requirement. How did you price the second redesign in the end?
That’s a good point. I actually didn’t do the full redesign. I talked it through with the client and made a few changes instead. 😅 Definitely learned to align expectations earlier!