Over the past few weeks, I’ve tried a bunch of AI models and subscriptions (Codex, Claude, GLM Coding Plan, OpenCode Go, and Command Code Go).
Here’s how it went.
By the time you read this, new models have probably come out and a lot has changed. Take it with a grain of salt.
Models
Model quality is important, but also how much work a subscription gets me before I hit those 5h or weekly quotas.
The harness is equally important, the less babysitting the better.
GPT models (via ChatGPT Pro 100)
- GPT 5.6 Sol was good, but burned through quota like there was no tomorrow.
- GPT 6 Sol improved on that quota burning.
- GPT 6.1 Sol is much more efficient. It’s probably the best balance a GPT model has had, although Opus 5.5 is still king of efficiency.
GPT 5.6 and 6 Luna are great cheap workers, but they can be slow, especially on Max (like go get coffee slow). I’d recommend High or xHigh for most tasks.
Astra is great, but doesn’t get you great mileage on a $100 subscription.
If you code for more than a few hours a day, I’d skip Codex’s $20 plan and go for $100. You could buy extra credits, but I’ve heard they’re not great value for money.
Claude / Opus models (via Claude Code)
Opus 4.x was decent, but the quota was tight.
Opus 5 came with better quota, but was dumb as a rock. One of the worst models I’ve used, and it talks like an alien.
Opus 5.5 is something else. It’s freaking smart, and I find Claude’s $100 subscription the most generous of the plans I’ve tried.
GLM models (via Z.ai)
I started with the $18/month GLM Coding Plan and eventually upgraded to $80. The quotas and resets are generous, and the models are decent. GLM used to be competitive with GPT 5.6 Sol, but not against Opus 5.5 and GPT 6.1 Sol anymore.
Buuut the biggest dealbreaker is the rate limiting. I couldn’t run more than a few agents at the same time, an issue I never had on Claude or Codex.
If you like GLM 5.3 Flash, you’re probably better off using it through OpenRouter.
OpenCode Go and Command Code
OpenCode Go ($10) is a decent backup when you run out of Codex quota. Amazing value. I wouldn’t use it as my daily driver. Also, the rate limiting is frustrating.
Command Code ($1 and $10 plans) offers amazing value on paper, but the throttling kills it for me too.
OpenRouter: DeepSeek 4.1 Flash
The model I’ve used most through OpenRouter is DeepSeek 4.1 Flash. I’d put it in the Luna class for intelligence, but it’s fast and cheap. Luna is still cheaper, though.
I’m willing to pay a little more for a model that keeps me moving fast.
Grok 4.6 and 4.7
Grok 4.6 and Grok 4.7 are decent. A bit expensive through the API. Can feel slow, maybe it’s me.
Other models I’ve tried
Also benchmarked latest models from Qwen, Hy, MiMo, Kimi, and MiniMax. None impressed enough to recommend.
I’m still curious about Meta Muse 1.3, but haven’t gotten around to trying it yet.
Harnesses
Codex vs. Claude Code
Claude Code for my job. Codex for personal experiments. I use both, and each has its perks and quirks. When I’m experimenting, I don’t wanna babysit. Codex is more autonomous and requires less attention.
I’m a desktop app person. I tried TUIs, but I keep coming back to desktop apps. Pi is the exception as it lets me try models from OpenRouter.
Pi
Still wrapping my head around Pi. Pretty vanilla out of the box, but you can customize the hell out of it.
First things I had built: a model router powered by Jev and an auto-review mode, also using Jev. Took some tuning, but both work well now.
I might move the decision-making to OpenAI’s Decisions API now that it’s out.
Ready to make Pi my daily driver? Not yet. Codex is still my go-to.
ZCode
I started testing ZCode with the GLM Coding Plan, since it’s the official harness. Looks a lot like Codex. Thread organization is better, though.
It connects to OpenCode Go, OpenRouter, and most API providers. I’ve mostly used it with GLM, OpenCode Go, and OpenRouter.
Workflows is one feature I really like. Similar to Claude’s dynamic workflows, but I think ZCode’s version looks better.
Stable, frequent updates. Still not as mature as Codex or Claude Code, but I’d take it over Amp, OpenCode, Kilo, or Command Code.
Benchmarks
Terminal-Bench 4 is the one I follow most these days. CursorBench used to be useful, but I’ve paid less attention since OpenAI and Cursor went through their little divorce.
New model drops? I check Artificial Analysis. Intelligence, cost, speed. All three.
A model being cost-effective doesn’t mean much if it’s slow as hell. (Luna Max, I’m looking at you.)
How I’m managing agents and projects
Lately, fewer threads per project. One main thread driving multiple subagents. I’m still experimenting, but it’s somewhat like Claude’s new Projects feature. Haven’t gotten access to that yet.
For personal experiments where I don’t need PRs, I don’t bother with worktrees.
Work is a different story. I often spin multiple repos, so I bring the relevant worktrees into one workspace. Keeps related work together without mixing it into everything else.
How I’d spend $200 a month
$100 on Codex or Claude. For heavy coding, those are the two I’d consider. I prefer Codex’s harness, but Claude’s Opus 5.5 quota is hard to beat. Pick whichever suits your workflow.
Want image generation, your own 24/7 dot, health data integration, and occasional resets too? I’d go with ChatGPT.
The other $100 stays with OpenRouter. Try new models, compare model quality, cost, and speed. Pay for inference when I need it.