A year of model-hopping

I've lost count of how many models I've run through this year. Some frontier, some open-source out of China. The honest summary is that no single one wins everything, so I stopped looking for the one and started matching models to jobs. This is the map I've ended up with.

The frontier rotation

My daily driver for coding is ChatGPT - mostly the GPT 5.5 and 5.4 models. That's where I do the bulk of the actual building, and it's the one I trust to stay coherent across a long agentic session without drifting.

I didn't start the year there. Around December into January I was leaning on Anthropic's Claude, the 4.5 through 4.6 models, before eventually switching back to GPT for the coding loop. Worth being precise about how I was reaching Claude, though: most of that wasn't through Anthropic directly, it was through Antigravity, Google's own setup. So my "Claude phase" was really an Antigravity phase that happened to be serving Claude.

Then there's Gemini. I tried the 3.1 Pro series fairly hard - 3.0 Pro, 3.1 Flash, the lineup. My take is split: they're genuinely good at frontend work, clean markup, sensible component structure, nice visual instincts. But they fall apart at being agentic. The moment a task needs to plan, hold state across tool calls, and grind through a multi-step loop, they lose the thread.

How the work actually splits

So the division of labor settled into something like this:

• GPT for coding and anything agentic. It's the workhorse.
• Opus or Claude for frontend, in the same slot Gemini fills - they have that same eye for UI that GPT doesn't quite match.

That's the frontier setup when I have full access to everything. The interesting part is what happens when I don't.

The open-source bench

For the minor stuff inside my agents - the supporting roles, not the headline work - I lean on open-source models, mostly out of China. Kimi, GLM, DeepSeek, and MiMo each have a seat. Right now that's Kimi 2.6, GLM 5.0 and 5.1, DeepSeek V4 Pro and V4 Flash, and MiMo 2.5 Pro.

Two of them have earned specific jobs:

When I don't have access to Opus or Claude, Kimi is my frontend fallback. It's the open-source model that gets closest to the frontier feel for UI work, so it slots right into the gap.

And for cybersecurity work I reach for MiMo. It's the one I trust for that domain, the same way I trust GPT for the general coding loop.

What the year taught me

The lesson isn't "model X is best." It's that the question is malformed. Best at what? GPT is best at holding a long agentic session together. Claude, Opus, and Gemini are best at frontend taste. Kimi is the best stand-in when the frontier ones are out of reach. MiMo is the one I want on security.

Treating models as interchangeable general-purpose brains is how you end up disappointed. Treating them as specialists with sharp edges - and routing work to the right edge - is how the whole setup starts to feel less like gambling and more like a toolkit.