Claude and Codex/GPT are available in many places now: an API, a CLI, another editor, or a chat window. Why I use Cursor instead of driving those models directly comes down to one line: the model is only half of the product, and the harness around it is the other half.
This is not an argument against Claude or Codex. I use them too. The question is where the same models produce the most useful work.
The harness I don’t have to build
The harness is the part of the product I would otherwise have to build myself. By “harness” I mean the system around the model:
- system prompts and context management
- tool design and retrieval
- guardrails that keep the agent from making a mess of the repository
A capable model still needs to know which files matter, how to inspect a project, and when to edit. Those choices change the result as much as the model name does. Cursor has a research team engineering this system full-time, so I open a project, describe the task, and let the agent inspect code, plan, and edit files where I already work.
Driving a raw model through an API or CLI means I own all of that myself:
- deciding which source files to send
- tracking the context window
- building and wiring tool calls
- maintaining the setup when the project changes
Some developers enjoy that work, and for special workflows it can be worth it. But it is extra engineering before the actual engineering starts.
I see the difference most clearly in frontend work, because I have basically zero frontend design skill. When I ask Cursor to build a web page, I consistently get a clean, modern Tailwind-based UI without specifying any of that. That default quality comes from the harness, not the raw model.
The harness also makes iteration less painful. I ask for a change, inspect the diff, point at a problem in the existing code, and keep moving without copying code between a terminal and a separate chat. The feedback loop stays close to the files, diagnostics, Git changes, and terminal output.
Composer and Grok: frontier-adjacent output at a fraction of the cost
Cursor’s first-party models are the strongest financial argument for the editor. Composer 2.5 is Cursor’s own agentic coding model, optimized for long-horizon, multi-file work inside the editor. It is my fast default workhorse for everyday agentic coding: understand a feature, trace the files involved, implement a focused change.
The numbers explain why. On the Artificial Analysis Coding Agent Index, Composer 2.5 scored 62 and ranked third overall, behind Claude Opus 4.7 max in Claude Code (66, about $4.10 per task) and GPT-5.5 xhigh in Codex (65, about $4.82 per task). Composer 2.5 Fast runs about $0.44 per task and standard about $0.07, which makes those two leaders roughly 10–60 times more expensive per task.
The quality gap can be within noise. On SWE-bench Multilingual, Composer 2.5 reached 79.8% versus 80.1% for Opus 4.7. One benchmark does not settle every real task, but it is why I reach for Composer first instead of assuming the most expensive option is always necessary.
Here is the division of labor I use. Prices and rankings below are as of July 2026, and they change fast.
| Model | What it is best at | Cost signal |
|---|---|---|
| Composer 2.5 | Fast everyday multi-file agentic coding | Standard: $0.50/$2.50 per M input/output tokens; Fast: $3/$15 |
| Grok 4.5 | Long, difficult multistep tasks beyond pure coding | $2/$6 per M input/output tokens; Fast: $4/$18 |
| Opus 4.7 max / GPT-5.5 xhigh | Top two entries in the cited Coding Agent Index | About $4.10 / $4.82 per task |
For harder work, I switch to Grok 4.5. Cursor describes it as its strongest model, trained jointly with SpaceXAI and built for long multistep tasks beyond pure coding. The announcement says it solves multistep tasks in under half the steps of comparable frontier models, which makes it my heavier option when a task has many dependent decisions.
The listed token prices are not always what I actually pay. Both models draw from the included first-party model pool on paid plans, so day-to-day usage often does not hit per-token billing at all. I can pick a capable model for the task without treating every prompt as a small purchasing decision.
Cheaper is not the same as best for every task. A difficult bug may still deserve a frontier model, and I still review diffs, run tests, and use judgment. For my normal development loop, though, Composer as the affordable default plus Grok as the escalation path beats paying the highest price by default.
It’s still just VS Code (and that matters more than it sounds)
Cursor being a VS Code fork matters more than it sounds: near-zero switching cost in, and near-zero switching cost out. My extensions, themes, keybindings, and settings carried over in minutes through settings import. My workflow muscle memory did not have to change at all.
I also inherit decades of editor maturity for free:
- language support and debuggers
- an integrated terminal and a Git UI
- remote and SSH development
None of that is exciting to list on a marketing page, but it is the infrastructure I depend on every day. The AI is added to an editor I already know, instead of a new tool I have to learn for the AI’s sake.
The exit is just as cheap as the entry. If Cursor stops being worth the subscription, moving back to VS Code costs nothing, because the skills and setup stay compatible. That asymmetry makes the subscription an easy bet.
Comparing tools by model names alone misses this. My day is spent navigating code, running the terminal, reviewing Git changes, and debugging. Cursor puts the model inside that routine instead of beside it.
Why I use Cursor: the short version
In AI coding, the bottleneck is no longer the model; it is the harness. So why I use Cursor is simple: it is where the same models, plus cheaper first-party ones, do the most work per dollar and per keystroke. I could still use Claude or Codex directly, and sometimes I do.
One honest caveat: prices and model rankings shift monthly, so check the current docs before deciding. My reason for staying is more stable than the numbers. I do not want to build and maintain my own coding-agent harness when I could spend that time shipping.