So here's the rundown of coding agents I've worked with this month.
Claude Code (the new build) - Sonnet 4.5 works beautifully, the new limits with the week of downtime make the app impractical and useless for real work. I wasted my money. It's not worth it.
Anthropic is not trustworthy at this point in the process.
Cursor - It's gotten better, use caps higher than Anthropic, but still overpriced. No five hour or weekly caps, but it's easy to blow through $60 of time. $200 will only get you a week of full time dev time.
Auto mode is terrible. Can't be used for serious super computing apps or agent frameworks. Probably workable for crud projects though.
Minimax 2 - As good or better than Claude Sonnet 4.5, but with a lot more enthusiasm. Reminds me of Gemma3 more than Claude.
Finding affordable hosting that won't break the bank for this one is going to be an ongoing problem. Even at Chutes which has the best pricing, you're still looking at about $300ish a month to run it for eight hours a day.
GLM 4.6 - Outstanding agent for simple apps and GUI. Comparable to Claude Sonnet 4.5, on simple single 3 layer crud apps. The model can't handle anything with even the slightest amount of complexity before bombing in extravagant ways. Loops, language switching, not pretty.
GPT-oss-20 and 120b - This model is damned weird. It's half way between o3 and GPT5, but in none of the ways I want it to be. Results are competent, but mid, adhocs way too often. Pricing: Same problem as minimax. 20b Loads locally on a 4070, but with severe context window limitations.
Jules2 - Meh. It's a Gemini agent. Workflow is interesting, but it lies.
Gemini CLI - This app is in a state of decay. It's worse than it's ever been. Heavy on refusals and emotional manipulation.
Build it with Gemini - Really good for react apps. Tell it to do anything other than a react app, it will take the "let's not and say we did" tact. When you point this out, it'll apologize, tell you that you're absolutely right, and keep building the wrong thing.
I'm holding on to my Gemini subscription because I'm hoping that Gemini Pro3 kills it. We'll see. Also, I like all the ancillary google services in the bundle.
The daily driver
Sticking with Codex this month as the daily driver.