Picking the Right Model
The best AI models change constantly. Check Artificial Analysis for the latest rankings. What I write here is a snapshot, it will age very quickly.
As of May 2026, the landscape looks like this:
- OpenAI makes the better coding models
- Claude makes the better UI design models
- Gemini is usable for design but weak at coding and agentic tasks
- Open Models are 90% there, but just not cutting edge
My daily driver for coding is GPT 5.5. It’s fast, smart, and token-efficient. For anything design-related — UI components, layouts, styling — I switch to Claude 4.6.
Note: Opus 4.7 has been disappointing. I still use 4.6.
Open Models
Models out of China like Kimi K2.6, GLM 5.1, and Deepseek v4 Pro are genuinely good. I use them for lighter work — CLI tasks, setting up VPS servers, quick scripts.
For serious projects, I stick with the cutting-edge closed models from the US. AI slop is still a thing. The best way to minimize it is to use the best available models.
Don’t bother coding with small local models at 40B parameters or less. The context window is too small to be useful for real agentic work.
Pricing: Tokens ≠ Cost
Open models are generally cheaper per token via API. But cheaper per token doesn’t always mean cheaper to run.
Models vary wildly in how many tokens they use for the same task. In one test, Deepseek v4 Flash used 240B tokens while GPT 5.5 used only 75B tokens — with the same result.
If a model is half the price per token but uses 3× more tokens, the “cheap” model is actually more expensive. Always check both the price per token and how many tokens a model uses per task.
GPT 5.5 is currently very token-efficient. More tokens don’t always mean more intelligence.
Check this benchmark comparison for the full picture.
Subscriptions vs API (Pay-per-Use)
AI providers are subsidizing subscriptions to gain market share. Both OpenAI and Claude offer these tiers:
- $20/month — entry level
- $100/month — pro
- $200/month — power user
With subscriptions you get session limits (5 hours) and weekly limits (7 days). But here’s the thing: if you hit those limits consistently, you can get roughly $2,000 worth of compute out of a $200 subscription.
These subsidies won’t last forever. But right now, they matter.
Ask yourself: do you want to pay $200/month for a subscription — or $2,000/month via API if you use a model outside OpenAI and Claude?
My Current Setup
- 2× $20 ChatGPT subscriptions — GPT 5.5 is my daily driver
- 1× $100 Claude subscription — only for design tasks, since GPT 5.5 is weak at UI
For the UI layer, I use T3 Code. You can add multiple subscriptions and use it with Claude, Codex, OpenCode, and Cursor. It has sidebar tabs like a browser. You can run 10 agents in parallel and switch between them.
I’ve used CLI tools before and enjoyed them. But a proper UI is just better for multi-agent work.
Why I Stopped Using Cursor
Cursor is a great tool. It has lots of features — code reviews, multi-model support, and more. You can pick any AI model you want. According to benchmarks, Claude Opus even produces better output in Cursor than in Claude Code, because the harness is better optimized.
But I have one problem with it:
Cursor can’t subsidize GPT 5.5 and Opus the way OpenAI and Claude do. Your $20/month plan burns through in just a few days. I easily spent $100/month while not even coding full-time.
Open Source Alternatives
If you like open source, check out OpenCode. It’s an open source coding tool similar to Codex and Claude Code. You can connect any model provider.
They also have their own provider called OpenCode Zen with two options:
- $10/month “Go” plan — access to open models like Kimi K2.6, MiniMax M2.7, etc. Good if you code occasionally, not all day. opencode.ai/go
- Pay-per-usage — also available with closed models like GPT and Claude
You can also connect OpenRouter as a provider. It has basically every model that exists. Want to try coding with a tiny 1B parameter model? You can. It won’t work, but it’s possible.
Any provider works. If you prefer European providers like Scaleway, Mistral, or Inceptron, you can connect those too.
Interestingly Codex is also open source. While Claude Code is closed source, though their code did accidentally leak early 2026.
Data Privacy
Regardless of which provider or tool you use: always change your settings to opt out of data sharing for model training. Every tool has this option somewhere. Use it.
