The AI Coding Harness I Settled On

I have $40 a month for AI coding. That’s the whole budget — I’m a student, not a startup with a credit line. So when “just throw more tokens at it” is the default advice in 2026, I had to actually figure out which harness earns its share of that $40.

This is the story of how I got there.

The shift nobody asked permission for

Coding in 2026 doesn’t look like coding in 2023. We went from writing every line by hand to agentic coding — what people now call “agentic engineering,” and what started as Andrej Karpathy’s “vibe coding.” Big companies and individual developers have already adopted it, because a good agent can leverage your output several times over.

The catch: you still have to review what the agent writes. Skip the review and you’re just front-loading technical debt onto future-you. Agentic coding moves the work, it doesn’t delete it.

The part the demos skip: it costs money

Every one of those leveraged hours burns tokens, and tokens aren’t free. That’s the constraint a student feels first. So I went looking for the harness that fit both my workflow and my wallet.

I tried a lot of them: Codex, Claude Code, OpenCode, Pi, Antigravity, and Cursor.

What I cut, and why

I ended up on Codex, Claude Code, and Pi now and then. The reasoning is simpler than it sounds: I only want frontier-lab models.

• OpenCode — I leaned on it, OpenCode Go especially. It offers a buffet of open-source models and the token economics are genuinely great. The output wasn’t. I’d hand it a plan already written out in a .md file, and it would still skip two or three things on the list — just not implement them. Debugging was worse: where a frontier model usually one-shots the fix, the open-source models took me two or three prompts to get there. Cheap to run, but “cheap” stops being the point when you’re paying for it in re-prompts.
• Antigravity — who’s still reaching for Gemini in 2026? (Kidding. Mostly.)
• Cursor — honestly the best harness of the bunch. Run Codex or Claude Code inside Cursor and you often get better results than running them on their own platforms. And their new Composer 2.5 model is fast.

So if Cursor’s harness is the best, why didn’t I just pick Cursor?

It came down to the limit type

Cursor only has a monthly limit, and the plan is $20. I could pour my whole $40 into Cursor, but it’s not worth it: once you burn through the $20 of included usage, it’s gone, and you’re either locked out until the calendar resets or pushed onto on-demand pricing — pay as you go. You can cap that at $20 if you want, but capped pay-as-you-go still doesn’t stretch the way Codex and Claude Code do.

So I split the budget: Codex + Claude Code, $20 + $20.

Frontier models give me output I’m actually happy to ship after review. And the limit structure is the real edge — both run on a 5-hour rolling window plus a weekly cap, not one hard monthly wall. Hit a wall at noon and you’re usually back in a few hours instead of waiting two weeks.

Does $20 each mean unlimited coding? No. I hit the limits constantly. Midterms were the worst — I was building the app for one of my exams and coding so much that I burned straight through Claude Code’s weekly cap with the reset still two days out. Codex is friendlier on limits, but it’s not bulletproof either: I once hit its weekly cap three days before reset, and that one was my own fault — too much browser use and too much xhigh reasoning, neither of which a $20 plan is built to absorb.

So no, the budget doesn’t work on its own. It works if you manage it well.

Where Pi fits

Pi isn’t a third model in my stack. It’s how I run Codex: a harness around the same frontier model, with room to load extensions and tighten the workflow around how I actually work.

Same frontier model underneath, better tooling wrapped around it. That’s the pattern across all of this: the model matters, but the harness around it decides whether the model is actually useful to you.

What I settled on

So that is where $40 a month landed me: Codex and Claude Code for the models, Pi when Codex needs a better shell. Not the flashiest stack. The one that still lets me ship after review, without pretending tokens are free.