Three layers, untangled
Almost everyone's confusion here comes from mixing up three different things. Once you separate them, the whole landscape gets simple. There's the room you work in, the assistant doing the work, and the bill.
The tricky part: some editors come with an agent baked in (Cursor, Antigravity), while an agent like Claude Code can run on its own in a terminal or drop into almost any editor. So the lines blur — but the three-layer model is still the right way to think about it. We'll take them one at a time.
Layer 1 · The editor (where you work)
All three below are close cousins — Cursor and Antigravity are both built on top of VS Code. The difference is how much AI is wired into the experience by default.
VS Code
MicrosoftOn its own, VS Code isn't an AI tool — it's a plain (excellent) code editor. Its superpower is that you can bolt any AI agent onto it with an extension: GitHub Copilot, Claude Code, Codex, Gemini, and more. Both Cursor and Antigravity are VS Code with AI built in, so this is the family they descend from.
Choose it when: you want a free, flexible base and you'll add the specific AI agent you prefer (very common — many people run Claude Code right inside VS Code).
Cursor
Anysphere · cursor.comCursor takes the familiar VS Code and puts AI everywhere: chat about your code, multi-file edits, and an "agent mode" that can make changes across your whole project. It's codebase-aware, polished, and the default many people reach for. A nice touch: you choose which model runs under the hood — Claude, GPT, or Gemini.
Choose it when: you want one tidy app that does it all and lets you swap the AI brain whenever you like.
Watch for: pricing is credit-based — heavy use can draw down your monthly allowance faster than you expect.
Antigravity
Google · antigravity.googleAntigravity (launched late 2025, expanded in 2026) flips the model: instead of helping you type, it gives you a control panel for setting AI agents loose on tasks while you supervise and approve. It's powered by Google's Gemini but can also run other models, including Claude.
Choose it when: you're deep in Google's world (Firebase, Cloud, Android) or you want to experiment with the newer "manage a team of agents" style of working.
Watch for: it's the youngest of the three, changes month to month, and can be resource-hungry on your computer.
Layer 2 · The AI agent (who does the typing)
This is the actual intelligence. An "agent" doesn't just suggest a line — it reads your project, plans, makes changes across many files, runs things, and fixes what broke. The big three each come from one of the major AI labs.
Claude Code
AnthropicClaude Code drives real, multi-step changes across an entire project and is widely regarded as one of the strongest agents for serious building. It's included with a Claude Pro or Max subscription (the same plan also gives you Claude on the web and desktop).
The honest catch: on the entry $20 plan, anything ambitious will bump into usage limits fairly often — more on this in the money section, because it's the single most important thing to plan for.
Codex
OpenAICodex is OpenAI's equivalent and is also top-tier for real coding work. It's bundled into ChatGPT subscriptions, so if your clients already pay for ChatGPT Plus, they likely already have it. The $20 Plus plan gets a decent amount done before limits; heavier users have higher tiers to grow into.
Nice if: you're already in the ChatGPT ecosystem and want coding power without a second subscription.
Gemini
GoogleGoogle's coding agent now lives inside Antigravity (its older standalone command-line tool is being retired in 2026). Its standout advantage is the Google ecosystem: tight, native connections to Workspace, Firebase, Google Cloud, Android, and AI Studio. If a client lives in Google's tools, that integration is real and valuable.
Nice if: your work is wrapped up in Google services and you want everything to talk to each other.
For raw coding power, the general consensus right now puts Claude Code and Codex ahead, with Gemini's edge being its ecosystem and speed. But this is the fastest-moving corner of the whole industry — leadership has flipped several times, and any of them can leap ahead with a single model release. Treat "which is best" as a this-month answer, not a permanent truth. The good news: switching is cheap, so it's fine to try one and change your mind.
How editors and agents combine
Because the layers mix, here's the plain-English version of the common setups. None is "correct" — they're just different combinations of workshop + assistant.
| Setup | What it feels like | Good for |
|---|---|---|
| Cursor (editor + its agent) | One polished app; pick Claude/GPT/Gemini inside it | Most people who want it all in one place |
| VS Code + Claude Code | Familiar free editor, with Claude's agent in the terminal beside it | People who want max coding power and don't mind the terminal |
| VS Code + Codex | Same idea, using OpenAI's agent | Already paying for ChatGPT |
| Antigravity (editor + Gemini) | A control panel for supervising AI agents | Google-ecosystem folks, agent experimenters |
Pick one: either Cursor (everything in one app, swap models freely) or VS Code + Claude Code (more power, slightly more terminal). Both cost $20/month to start. Don't agonize — you can change later in an afternoon.
Layer 3 · The money (and where you hit the wall)
Here's the part nobody tells beginners clearly: the free tiers are test drives, not enough to actually build with. To do real work, you need a paid plan — and the $20 tier is the real entry point for all of them.
Where the $20 plans run out
This is the most important practical thing to understand, and it confirms what trips people up most:
- Claude Code on the $20 (Pro) plan: usage refills on a rolling few-hour window. It's great for learning and small builds, but anything ambitious will bump into the limit relatively often — you'll get told to wait until the window resets. If your clients try to build something big on Pro, expect this.
- Codex on the $20 (ChatGPT Plus) plan: can get a decent amount done before limits in each window — a comfortable starting point for most people.
- Cursor on the $20 (Pro) plan: unlimited routine assistance plus a monthly pool of credits for the most powerful models; the credit pool is what runs low under heavy use.
Don't start on a $100–$200 plan "just in case." Start at $20, and upgrade only after you've been throttled mid-build for a week or two. That's the real signal. If a client gets seriously into building, the jump to a $100 tier (5× the usage) is usually the sweet spot before the $200 power-user tier.
| Plan | Entry $20 tier | Next step up |
|---|---|---|
| Claude (Pro / Max) | Pro — bumps limits often on big builds | Max 5× ($100) → Max 20× ($200) |
| ChatGPT (Plus / Pro) | Plus — decent amount per window | Pro ($100, ~5× Codex) → ($200) |
| Cursor (Pro / Pro+ / Ultra) | Pro — $20 credit pool + unlimited routine | Pro+ ($60) → Ultra ($200) |
Figures are a 2026 snapshot; all three change pricing often, so confirm on each company's pricing page.
API & pay-per-use — and why to mostly avoid it
You'll see another way to pay: the API, where instead of a flat monthly fee you pay per "token" (roughly, per chunk of text the AI reads and writes). For vibecoding, this is usually the wrong choice — here's why in plain terms.
- It's pay-as-you-go. No subscription; you buy credits and each request draws them down. Great for powering an app you're building, but expensive for your own long coding sessions.
- Coding burns tokens fast. An agent re-reading your whole project on every step adds up quickly. People who code all day on the API can spend many hundreds — sometimes thousands — of dollars a month, versus a flat $100–$200 subscription that covers the same work.
- Bills are unpredictable. A flat plan is a known number each month. The API is not, and surprise invoices are a real, documented thing.
Stay on a flat monthly subscription. If your tool ever offers to "add API credits" to keep going after you hit a limit, it's usually cheaper to simply wait for your window to reset or move up one subscription tier instead. Save the API for when you're building a product that calls the AI on its own — not for your personal coding time.
What a real session looks like
To make the limits concrete, here's a realistic Saturday for a beginner building a small dashboard with Claude Code on the $20 plan.
EXAMPLE A morning of building on the $20 tier
A clean starter setup we'd recommend
- Pick one combo: Cursor (Pro, $20) or VS Code + Claude Code (Pro, $20).
- Connect your GitHub (and keep repos Private — see the hosting guide).
- Stay on the flat subscription; decline API/pay-per-use prompts.
- Work in focused sessions; if you're constantly throttled, upgrade to a $100 tier.
- Don't subscribe to all three labs. One agent at $20 is plenty to start; add or switch only if you have a reason.
Where this fits in the series
These tools are the step up from the app builders in the platforms guide. The natural path: rough out an idea in Lovable or Bolt, then move into a real editor with an agent — Cursor or VS Code + Claude Code — to refine, secure, and extend it. From there, the hosting guide takes over for putting it online.
Prototype in an app builder → graduate to an editor + AI agent (this guide) → host it with the right stack (hosting guide). Three guides, one path from idea to live app.