Everyone’s asked me the same question for a month now: you quit Claude Code... so what do you use instead?
Thirty days in, I have a real answer. It comes in two parts, and the second part is the one that saves you the money.
But the question underneath it deserves a careful answer too: is there a setup cheaper than Claude Code that still does the job? For my business, yes — with one big condition we’ll get to.
The condition is which person you are this quarter, not which tool you use.
Claude Code Is Two Things — Replace Them Separately
Part one: a brain. Claude Code ships with Anthropic’s models inside, but intelligence is a market now. I run a Venice API key. (I work for Venice, bias declared. The video this post comes from wasn’t sponsored; I was making Venice videos before they ever hired me.) You could use OpenRouter, Requesty, PPQ, NanoGPT, or go straight through Kimi or DeepSeek.
Part two: everything else. Claude Code is more than intelligence. You can chat with intelligence in a chat box. Claude Code is agentic: it does things in your files. So you need a harness: a runtime environment, a sandbox, where the AI can work.
Mine is OpenCode, an open-source agent app. Agent Zero and DeepSeek’s V4 Flash live in the same open lane. Install is one pasted line in a terminal, and there’s a desktop app now.
Provider, harness, picked separately. I already ran a real client job on this exact stack (the whole afternoon cost about $9) and wrote up the full test with receipts.
What the $200 Was Actually Buying
Subsidized inference: a flat subscription covering more token burn than the tokens cost. That’s my read, anyway, and the walls are moving: limits are going down. These companies have to figure out profitability eventually. I’d rather get ahead of that than be surprised by it.
The subscription also hides what you burn. On the flat plan I caught myself thinking in fantasy math: I could use 500 million tokens in a month... in a day... within my five-hour limit. Twice. That’s not reality. That’s a dreamland.
When you pay per call, you notice consumption inside a week. It’s the difference between a buffet and a metered tab.
If you want the baseline first, I broke down what Claude Code costs plan by plan. And the full story of why I walked is in why I cancelled my $200/month subscription.
The Failure First: Doing More, Shipping Less
My honest confession from those subsidized months: I was doing more, shipping less, getting less done, distracted by all the shiny objects I could create with all those essentially free tokens.
I have a little graveyard of side projects I started because I had the tokens to try them. (Why not, right?) Every one of them was a distraction from the goals.
Hormozi said it recently and it stung: doing more with AI doesn’t mean you’re moving your business forward.
What I landed on: I don’t need all the extra tokens. I need me — then I need AI to automate the labor. That order matters.
Person A, Person B: Which One Are You This Quarter?
Person A runs a business. She builds occasionally, but most of her AI usage is automated workflows: repurposing content, drafting client reports, running the admin. She needs good intelligence flowing through a system she already built. Her month can run far under $200.
Person B is in a building phase. My day job right now is building an AI editor, genuinely hard cutting-edge work, and for that I still reach for frontier models, the best I can get. Person B needs the tokens.
The point: they’re cycles, not identities. Sometimes you’re Person A. Sometimes you’re Person B. Segment by work type, and notice who pays for the inference — if an employer or client covers it, you’re in a different boat than the owner paying for every credit.
Which one are you this quarter? I honestly can’t answer that for you. But it’s the question that decides whether canceling is genius or self-sabotage.
The Workspace Is What Makes Cheap Models Work
My working rule: build with frontier, execute with open-weight. Use the best models to build the system, then let economical models run it, following state-of-the-art instructions.
On camera I demoed this live with GLM 5.2, a mid-priced open-weight model, inside VSCodium and OpenCode:
AGENTS.md who the agent is, how we talk, hard rules (under 40 lines) skills/ SKILL.md files — one repeatable job each agents/ reviewer.md — checks every draft against the rules
The test business was a one-person fitness coach. First skill: a monthly progress summary (what you did, what changed, what comes next, one line of encouragement). Test client Jill Jacobs got her report, the reviewer agent checked it against the voice rules, verdict: pass.
Then the compounding trick: I edit a draft, the agent diffs my version against its own, finds what I changed, and proposes one new rule, then waits for my approval before touching anything. On camera it banned em dashes forever, in AGENTS.md and the reviewer’s checklist. (A bummer, honestly. I love an em dash. But readers pattern-match it to AI slop now... so it’s gone.)
That whole structure is a harness: the instructions, skills, and reviewer files you own. It’s what lets a model that’s less intelligent accomplish the task faster, cheaper, probably just as good. I’m a year into my main workspace and I revamp it every quarter. It keeps getting better, like an employee you’re training.
Model Routing: The Savings Lever Nobody Thinks About
Model routing: every job goes to the cheapest model that can do it well.
My ladder: plan on Kimi K3 Fast (a powerful open-weight model at less than half the per-token price of Anthropic’s Fable), build on GLM 5.2, and send mechanical jobs (summarize this, fix that grammar, what does the site say right now) to a super-cheap model.
The cost ladder runs pennies, dollars, tens of dollars per million tokens. Claude Code was burning tokens on review passes I never noticed, because they were subsidized. I notice now, and routing is how I spend the noticing well.
What My Month Costs Now
Honest answer, 30 days in: I don’t have a clean invoice to show you yet. The truthful claim is theoretical: Person A months, where I mostly run what I’ve built, land far under $200. A Person B building month could match it.
But the meter tells the truth now, and that changes behavior more than any subscription price did. That’s the real answer to cheaper than Claude Code: it’s cheaper when you’re running, comparable when you’re building — and now you can see both.
The System Is the Savings
If your AI life right now is nine subscriptions, four browser tabs, and copy-paste between them, a cheaper tool fixes nothing. Juggling tools, no system. The tool was never the bottleneck.
What fixes it is the order you build in: harness first, skills next, reviewers, routing. Then systems run the business, and your time is yours. The cheaper bill follows on its own.
That’s what AI Captains Academy is for: a clear order to learn in, feedback on your real work, and a crew to think with. We start builders on Claude Code (bias declared; it’s the gentlest first mile), we teach exactly the workspace skills this post runs on, and the full Armory of pre-built harnesses is included at every tier: $83/month or $497/year, with a 7-day free trial.
Questions people actually ask
- What is a cheaper alternative to Claude Code?
- The stack I run is OpenCode, an open-source agent harness, pointed at pay-per-token intelligence: a Venice API key, OpenRouter, Requesty, or Kimi and DeepSeek direct. The intelligence and the harness are separate choices, and you can replace each without the other. My test of this stack on a real client job cost about $9 for the afternoon.
- How much cheaper is pay-per-token than a Claude subscription?
- The reference point is a $200/month subscription. Open-weight models cost pennies to tens of dollars per million tokens (Kimi K3 Fast, my planning model, runs at less than half the per-token price of Anthropic's Fable). A month of running established workflows can land far under $200; a heavy building phase can match it. Your work type decides.
- Should I cancel my Claude Code subscription?
- Ask which person you are this quarter. Person A mostly runs established workflows and can run them cheaper per token. Person B is in a building phase and still needs frontier models daily. They are phases, and you move between them as the work changes. If an employer or client pays for your inference, the math changes again. I cancelled, and I kept a way to reach frontier models when I build.
- Do open-weight models work as well as Claude?
- Unaided, no. The capability ceiling (the point where a cheaper model can't keep up) is real. Inside a good workspace (startup instructions, skill files, a reviewer agent that checks drafts), a cheaper model accomplishes the same task faster and probably just as well. The workspace closes most of the gap, which is why I build with frontier models and execute with open-weight ones.
- What is model routing?
- Model routing means sending each job to the cheapest model that can do it well. I plan on Kimi K3 Fast, a strong open-weight model; build on GLM 5.2, a mid-priced one; and send mechanical tasks like summaries and small fixes to a super-cheap model. The ladder runs from pennies to tens of dollars per million tokens.
- Do I need to be a developer to do this?
- No. I'm not a developer or an engineer. I'm a builder. The whole workspace is plain-text files: an AGENTS.md the agent reads on startup, skill files describing your processes, a reviewer profile. If you can describe how you want the work done, you can write the files, or have the agent interview you and draft them.
