State of AI for me May 20, 2026
-
Google just launched Gemini 3.5 Flash to mixed reviews. One of the main criticisms is it’s not as cheap as advertised and is much more expensive than 3.1 Flash. While true, I think they weren’t trying to replace 3.1 “Flash”. They were trying to replace 3.1 “Pro”, and I think they achieved that. I’ve been on Google’s ¥2900/month plan for the last month and have been on Claude’s and Codex’s paid plans before too. My assessment, even after different models have been coming out, is kind of unchanged and easy to explain. Not sure if you guys agree but I see the three like this…
Claude Code (Opus and Sonnet)– Best for starting completely new projects from ground zero. It understands and follows through on entire PRD’s the best of all three in this area. Opus 4.7 seems to go down rabbit holes though, often eating more cookies than I wanted to feed it.
Codex (GPT 5.4. Not sure about 5.5)– Absolutely the best for hunting down bugs efficiently. Usually one-shots the fixes. Although it seems efficient(don’t think I ever hit my daily limits), it’s slower to finish its work.
Gemini 3.1 Pro & Google Stitch – Most cost effective of all three. Best front-end designs period. Has Google search at its fingertips when you want it to check for options for feature building. As of today, 3.5 Flash is supposedly as good as 3.1 and is 4 times faster. I can verify the speed for sure. However, it seems to eat twice as many cookies as 3.1 for the same job. Price per token in/out, while cheaper, makes it a wash. You get the same job done by both 3.1 Pro and 3.5 Flash, roughly the same price, but I guess 3.5 Flash gives you extra speed.
Summary...The worst part of Gemini models is that the only way it can one-shot hard bug fixes is to give it pictures, and to explain the step-by-steps the user did to encounter the bug. Codex didn’t need any of that. I could just say the shit don’t work and it fixed it while I watched a YouTube video. Claude was decent at fixing shit but seemed to try too much other shit in the process. I mean, why was Claude attempting to start a game of Kred? It can’t do that without multiple players! Lol. Codex may be the most well-rounded of the three with regard to coding. It’s not great at design. It’s not great at keeping all of my intent on new projects. But in brick-by-brick mode(which includes bug fixes), it’s the best. Claude’s 5-hour limits are infuriating but the models are super smart(they seem to get and keep my intent no matter what). Gemini is the jack of all trades, in the sense that the pro plan gives you access to the entire Google AI ecosystem (Nano Banana, Stitch, Antigravity, NotebookLM) but needs refined/detailed prompts.
For now, I’m gonna go for one more month with Google then see how Anthropic is treating their customers. I might go back to them. Not sure I’ll go back to Open AI just yet. Fast bug fixes are great, but not worth it based on just that. One thing that I keep in mind is that I build games. I am not trying to solve complicated math or science problems. Sonnet, GPT 5.4 on medium, Gemini Flash have gotten shit done for me in the past. Cranking up to Opus, GPT 5.4 high and Gemini Pro models for shit in the weeds has worked out fine too, meaning I don’t necessarily need to pay for the model that out-benchmarks all the others. It’s not the models...it’s me and how I learn to use the models to their strengths.
Hello! It looks like you're interested in this conversation, but you don't have an account yet.
Getting fed up of having to scroll through the same posts each visit? When you register for an account, you'll always come back to exactly where you were before, and choose to be notified of new replies (either via email, or push notification). You'll also be able to save bookmarks and upvote posts to show your appreciation to other community members.
With your input, this post could be even better 💗
Register Login