Cole McIntosh

Founder @Mellow AI

|
|Resume

Every Max Coding Agent Plan, Tested

The $200 coding agent plan is the new default if you actually live in these tools. I paid for all four: Codex on ChatGPT Pro 20x, Cursor Ultra, Claude Code on Max 20x, and Factory Max. Same work. Shipping product, porting libraries, reviewing diffs, kicking off jobs I did not want to babysit.

The question is not which model is smartest this week. It is which surface makes the work disappear, and which one just moves the bottleneck onto you.

Codex

Codex is a worker you dispatch. CLI, IDE extension, web, cloud tasks, GitHub review. The $200 Pro 20x seat also buys the rest of ChatGPT, which matters if you already wanted that anyway.

It is strongest when the task is specified. Give it a ticket with a target and a test and it will come back with a PR. Cloud tasks keep running after you close the laptop. The review bot is useful at volume, which is the part most people undersell. The CLI is light. It does not try to be your editor.

Pros

  • Cloud tasks that survive your session
  • GitHub review that scales with output
  • One bill if you already live in ChatGPT
  • Fast on well-scoped implementation

Cons

  • The product is split across surfaces. You feel the seams
  • Locked to OpenAI models
  • Five-hour windows turn the day into a quota game
  • Underspecified work gets a confident invention, not a question

Keep Codex if you want an async implementer and you were going to pay for ChatGPT Pro regardless. Do not keep it if you expected a place to live.

Cursor

Cursor is the editor. Tab, inline edit, agent, cloud agents, review. Ultra is $200 for roughly 20x Pro usage and a $400 frontier-model pool. The real feature is not any one model. It is being able to pick.

I ported a Python library to TypeScript in an evening by running agents in parallel on separate branches and pointing Grok at the PRs. That only works if dispatch is cheap and the answers come back fast enough to stay in the loop. Cursor is built for that loop. You stay in the code. You see the diff. You redirect. You dispatch again.

Tab is still the most honest feature in the category. It removes keystrokes without creating a review queue.

Pros

  • One surface for the whole day
  • Model choice: speed when you want parallelism, judgment when you want a hard call
  • Cloud agents and parallel branches as a habit, not a treat
  • Tab that deletes work instead of relocating it

Cons

  • Credit accounting is a fog. You feel rich until you don't
  • Max mode will eat the budget if you treat it as default
  • Easy to generate more review than you can absorb
  • You paid for a very expensive editor if what you wanted was to leave the editor

Keep Cursor if the editor is already where you think. The risk is throughput without judgment. Four confident agents will happily answer a question nobody should have asked.

Claude Code

Claude Code is the sharpest instrument of the four. Terminal-native, now with IDE hooks, running on Anthropic's models. Max 20x is $200 for 20x Pro usage per session, shared with claude.ai.

It is the one I trust on a hard, existing codebase. Before it edits, it looks. Agentic search is the product: it walks the repo, follows the types, and comes back with a smaller diff than the others. The sessions stay coherent. Skills, hooks, and MCP make it feel like a tool you can shape instead of a chat you tolerate.

Pros

  • Best at understanding a repo before touching it
  • Highest-judgment diffs. Less plausible nonsense
  • Long sessions that do not lose the thread
  • Hooks, skills, and MCP if you want to build around it

Cons

  • Anthropic only. Want a different model, leave
  • Chat and coding share one pool. A long conversation burns a coding session
  • Weekly and five-hour limits still bite if you run it all day
  • Weak at fanning out a queue of tickets you do not want to watch

Keep Claude Code for the problem you cannot afford to fake. Pair it with something that does volume. Alone, it is a scalpel you will try to use as a factory.

Factory

Factory is not a pair programmer. It is a system for handing work to Droids. App, CLI, SDK, cloud and local background agents, Droid Computers, Missions. Max is $200 for about 10x Pro usage and early features. When the premium pool runs out, Droid Core open-weight models keep you moving.

The interesting bet is specialization and walk-away work. You assign a ticket. You come back to a PR. Model routing is built in, so you are not married to one lab. If your day is a queue of well-specified issues across repos, this is the only one of the four that feels designed for that sentence.

Pros

  • Built for delegation, not pairing
  • Model-agnostic, with a free open-weight fallback
  • Missions and remote computers for work that should not need you
  • Feels like an engineering system, not a chat window

Cons

  • More product to learn than the work sometimes deserves
  • Three rolling rate-limit windows plus Extra Usage. Missions want Extra Usage on
  • The interactive loop is looser than Claude Code
  • Easy to inherit a PR you do not understand
  • The SDLC platform is more than a solo day usually needs

Keep Factory if you can specify work well enough to not watch it. Do not keep it if you want to stay in the loop. Over-delegation is how you get a review tax that never goes away.

What I would actually pay for

If I could keep one: Cursor. I already live in the editor, and model choice matters more than any single model's peak week.

If I could keep two: Cursor and Claude Code. Editor for the day. Terminal for the problem that needs judgment.

Codex if the ChatGPT seat was happening anyway and I want an async implementer on the same bill.

Factory if the work is a queue I can specify and leave.

Do not stack all four. The tax is not the $800. The tax is four review queues. Writing the code got cheap. Deciding which code belongs did not. Buy the surface that matches how you work, then protect your attention like it is the scarce resource, because it is.