While Everyone Posts About Claude 4, I'm Using Gemini Flash

Everyone's talking about Claude 4 and new model releases. Meanwhile, I spent this evening experimenting with Google Gemini Flash.

I picked it almost by accident. Was supposed to use Pro, but I grabbed Flash instead. Normally I'd switch back, but I decided to let it cook for a few hours.

Here's what shocked me: it didn't break things. All my guardrails, memory banks, and system instructions I've built up actually work with Flash. For the last three hours, I've been coding an entire admin interface on top of a MobilePayment API in my solution. Flash didn't make significantly bigger mistakes than the expensive models, and it was dramatically faster.

Now I'm confused in the best way possible.

The model is 20x cheaper and noticeably faster, yet the output quality hasn't suffered. This breaks my entire mental model about AI-assisted development. I've been operating on the assumption that you go expensive first and accept slower iterations as the price of quality.

Not anymore, apparently.

The question this raises is bigger than just savings. If Gemini Flash can handle my complex vibe coding projects with the same reliability as Sonnet or Gemini Pro, then maybe the whole approach needs rethinking. Maybe I've been overpaying for capability I don't actually need.

Or maybe my guardrails are just that good at compensating for a less capable model.

I'm going to give Flash full access this weekend and actually push it hard. No switching back yet. Let's see if it holds up under real pressure or if I hit a wall eventually.

The implications if it does work: my cost per prompt drops dramatically, iterations speed up significantly, and I need to completely reconsider my spending on model selection.

Isn't it funny when the thing you thought you understood suddenly isn't as simple as you believed?