I switched from Cursor to Claude Code in August 2025 and never switched back. Nine months and several large codebases later, this is the unvarnished review of Anthropic’s terminal-native AI coding agent — what it does that nothing else does, where it still falls short, and whether the $20/month Claude Pro plan is actually enough.
Principaux enseignements
- Best in class at multi-file refactors and long-running agentic tasks (10–30 min runs without breaking).
- Terminal-native design beats IDE-bolted-on agents for anything that involves the shell, git, or build systems.
- MCP ecosystem (200+ servers) makes it the only agent that meaningfully integrates with your existing tools.
- $20/mo Claude Pro plan is enough for most devs; heavy users will hit limits and need the $200 Max tier or API.
- Loses to Cursor only on raw inline-completion speed and visual diff review.
What Claude Code actually is
A CLI tool you install with npm install -g @anthropic-ai/claude-code (or via the official installer) that gives you an interactive AI coding agent inside any terminal, in any directory, with full access to your shell, file system, and git history.
It is not an IDE plugin. It does not have a UI of its own beyond the terminal. It does not need an editor — though it integrates with VS Code, JetBrains, and Neovim if you want. You type natural language; it edits files, runs commands, reads logs, opens git diffs, and iterates until the task is done.
The model behind it is Claude Sonnet 4.7 by default (Opus 4.6 / 4.7 on the Max plan and via API). Anthropic ships major updates roughly monthly.
Pricing — and what each plan actually buys you
There are three ways to pay for Claude Code:
| Plan | Prix | What you get | Good for |
|---|---|---|---|
| Claude Pro | $20/mo | Sonnet 4.7 · ~40 hours of Code per week · 5-hour usage windows | Solo devs, side projects |
| Claude Max ($100) | $100/mo | 5× Pro limits · Opus 4.7 access · longer windows | Professional devs at small companies |
| Claude Max ($200) | $200/mo | 20× Pro limits · Opus by default · priority access | Heavy users / consultants |
| API pay-as-you-go | Per token | No subscription · prompt caching brings cost down ~90% | Teams, automation, very heavy use |
The Pro plan handles most independent work. I personally exceeded it within three months of daily use on a large monorepo and moved to Max $100, which has held up since. The API route is best when you’re integrating Claude Code into CI or running unattended agents — prompt caching makes it competitive on cost.
The five things Claude Code does better than anything else
1. Long-running agentic tasks that don’t fall apart
Give Claude Code a task like “find every place we call the old Stripe API and migrate to the new one, run tests after each change, and commit per logical unit” and walk away. A 20-minute run with 40 tool calls usually completes successfully. Cursor’s agent mode gives up sooner; Aider gets lost in larger codebases; Codex (OpenAI’s variant) works but burns through context faster.
The reason is structural: Claude Code maintains explicit task state, uses sub-agents to keep the main context lean, and aggressively compacts older turns. Anthropic clearly built this for hour-scale work, not five-prompt sessions.
2. Codebase awareness without indexing
Cursor builds a vector index of your repo on first open. That’s fast at retrieval but stale within hours and brittle on huge monorepos. Claude Code reads files on demand using glob, grep, and Read tools. There’s no index to keep fresh. On a 400k-LOC Go codebase I work on, Claude Code’s “find every place X is defined and audit it” is consistently more accurate than Cursor’s @Codebase because it’s actually reading current file contents, not querying an embedding.
The trade-off: it’s slightly slower per-query. For interactive work, Cursor wins. For correctness on big tasks, Claude Code wins.
3. MCP — the integration layer nothing else has
Model Context Protocol is Anthropic’s open standard for letting AI agents talk to external systems. As of mid-2026 there are ~200 official MCP servers (GitHub, Slack, Linear, Postgres, Sentry, BrowserBase, Playwright, AWS, Cloudflare, you name it) and a long tail of community ones.
In practice this means I can ask Claude Code “look at the failing tests from the last CI run on this PR, find the regression, and fix it” and it actually reaches into GitHub Actions, reads the logs, narrows the failure, and ships the patch — all in one session. That kind of cross-tool agency is functionally impossible with Cursor or Copilot today.
4. Hooks — deterministic safety on a non-deterministic agent
Claude Code lets you wire shell scripts to lifecycle events: PreToolUse, PostToolUse, UserPromptSubmit, Stop. The simplest valuable use: a PreToolUse hook that blocks file writes outside the current repo. The most valuable I’ve seen: a PostToolUse hook that runs your linter and formatter automatically, so the agent’s output is always clean.
This is the single most underrated feature. It lets you make Claude Code stricter than the model alone could ever be.
5. Custom skills + sub-agents
You can drop a markdown file in ~/.claude/skills/ and it becomes a callable capability — for instance, a release-notes skill that knows your project’s changelog format. Sub-agents (general-purpose, code-reviewer, etc.) let you delegate sub-tasks to ephemeral context windows so the parent agent stays focused. Together they let you scale Claude Code from “one prompt at a time” to “an actual workflow.”
The three things it still gets wrong
1. Inline completions don’t exist
If your daily flow is “type, get a grey suggestion, press Tab,” Claude Code is the wrong tool. Use Cursor or Copilot for that. Claude Code is an agent, not an autocomplete.
2. Visual diff review is weak
The TUI shows diffs as colored text. For a 600-line refactor that’s fine but tedious; Cursor’s side-by-side diff in the editor is genuinely better for review. I usually git diff in another pane or open the IDE side-by-side to review.
3. Rate limit windows are blunt
On the Pro plan, you’re throttled in 5-hour windows. If you burn through your quota in the morning, you’re locked out until afternoon. There’s no soft warning; you just get told to come back later. The Max plans mostly fix this but at meaningful price.
Claude Code vs the alternatives
| Capability | Claude Code | Cursor | GitHub Copilot | Aider |
|---|---|---|---|---|
| Long-running agent | Excellent (20+ min) | OK (5–10 min) | Limited | OK |
| Inline completions | None | Excellent | Excellent | None |
| Codebase awareness | On-demand, accurate | Indexed, fast | Limited | On-demand |
| MCP / integrations | 200+ servers | Limited | GitHub only | None |
| Hooks / safety | Yes (4 events) | Non | Non | Non |
| Visual diff UI | TUI only | Side-by-side | OK | TUI only |
| Best models | Claude 4.7 | GPT-5, Claude 4.7, Gemini 3 | GPT-5, Claude 4.7 | Any (BYOK) |
| Starting price | $20/mo | $20/mo | $10/mo | Free (BYOK) |
Pros and cons
Claude Code pros
- Best-in-class long-running agent reliability
- Terminal-native — fits any editor or none at all
- MCP ecosystem opens up dozens of integrations
- Hooks let you make it stricter than the model
- Sonnet 4.7 is currently the best coding model
- $20/mo Pro plan is genuinely usable
Claude Code cons
- No inline completions — wrong tool if you want Tab-Tab-Tab
- Visual diff review weaker than Cursor
- 5-hour rate limit windows on Pro can sting
- Steeper learning curve than IDE plugins
- API mode can get expensive without prompt caching
Who Claude Code is for — and who it isn’t
Buy Claude Code if:
– You work on real codebases (10k+ LOC) and want agentic refactors that hold together
– You live in the terminal and value being editor-agnostic
– You want integrations with GitHub, databases, browsers, etc. via MCP
– You’re willing to learn a more powerful but less hand-holdy tool
Don’t buy Claude Code if:
– You mostly want autocomplete while you type — get Copilot
– You prefer a polished GUI for everything — get Cursor
– Your work is mostly single-file prototyping — anything works
Performance after 9 months of real use
A rough log of what Claude Code has shipped for me, unattended, in production:
– Migrated a 12-service Node monorepo from CommonJS to ESM (3 hours, 1 supervised run)
– Audited a Rails 7 app for N+1 queries and fixed 31 of them with tests (~90 minutes)
– Ported a 4k-line React class-component bundle to function components + hooks (50 minutes)
– Wrote and shipped 28 PRs over a weekend that fixed papercut issues from a backlog
The failure modes I’ve seen:
– Got stuck in a tool-call loop once trying to use a flaky MCP server (~$3 wasted on the API). Killed it manually.
– Once committed a .env.local file because the gitignore had drifted. The PreToolUse hook I added afterwards prevents this class of issue.
– Sometimes over-edits when given vague instructions. Tighter prompts and the Plan agent help.
It’s not magic. But it’s the closest thing to a junior engineer who never gets tired.
FAQ
Is Claude Code free?
There’s a 14-day free trial of Claude Pro. After that, $20/mo Pro is the minimum tier. The CLI itself is open-source under MIT, but it requires a paid Anthropic account or API key to function.
Does Claude Code work without internet?
No. The CLI is the client; the model runs on Anthropic’s servers. You’ll need a connection. There’s no on-device mode and Anthropic hasn’t suggested one is coming.
Can I use Claude Code with my own model (BYOK)?
Not officially. The CLI is Claude-only. If you want a similar UX with BYOK, look at Aider or Cline, both of which support OpenAI / Anthropic / local models.
Is Claude Code better than Cursor?
For long agentic tasks and terminal/CLI work: clearly yes. For typing-with-autocomplete and visual editing: clearly no. Most serious devs end up using both — Cursor for fast iteration in the editor, Claude Code for the heavier “make this thing happen” tasks. They’re complementary, not competing.
Is Claude Code better than ChatGPT for coding?
For actually doing the work — editing files, running commands, finishing multi-step tasks in your real repo — yes, decisively. ChatGPT (even with the coding-focused GPT-5.5) is a chat interface: you copy code in and out. Claude Code is an agent that operates directly on your codebase, shell, and git. For one-off “explain this” or “write a function” questions, ChatGPT is fine and faster to reach. For shipping changes to a real project, Claude Code is a different category of tool.
Is Claude Code still the best coding agent in 2026?
For long-running, terminal-native agentic work, it’s the one to beat — the reliability on 20-minute multi-step tasks and the MCP integration ecosystem keep it ahead. Open-weight models like Kimi K2.6 and GLM-5.1 have closed the raw-capability gap and cost far less, and Cursor is better inside the editor. But for “give it a complex task and trust it to finish,” Claude Code remains our top pick.
Is Claude Code actually worth it?
If you work on real codebases daily: yes, easily — the $20/month Pro plan pays for itself the first time it completes a refactor or migration you’d have spent hours on. If you only code occasionally or mostly want autocomplete, it’s the wrong tool and a cheaper option (or Cursor) fits better. The honest test: if you have tasks you’d hand to a junior engineer, Claude Code is worth it.
What’s the difference between Pro and Max for Claude Code specifically?
Pro gives you Sonnet 4.7 with limited 5-hour windows (about 40 hours per week of intensive use). Max ($100) gives you 5× the limits plus Opus 4.7 access for harder tasks. Max ($200) gives you 20× and Opus by default. If you use Claude Code for 4+ hours a day, you’ll outgrow Pro within a month.
Is my code sent to Anthropic?
Yes — context you share with the agent is sent to Anthropic’s API for inference. Anthropic does not train on Pro/Max API customer data by default. For paranoid environments, the API offers Zero Data Retention agreements; on the consumer plans you get standard privacy terms.
Does it work with non-English codebases / comments?
Yes. Claude handles non-English content fine, including code with Arabic, Chinese, Russian, etc. comments and identifiers.
Résultat
Claude Code is the first AI coding tool that genuinely changed how I work, not just how fast I type. It is not the right tool if you want autocomplete; it is the right tool if you want an agent that can be given a complex task and trusted to finish it.
The $20/month Pro plan is one of the highest-value subscriptions in software development right now. The $100 Max plan is the right answer for anyone using it daily.
If you’ve tried it before and bounced off, the version shipping in mid-2026 (Sonnet 4.7 + the latest CLI) is genuinely a different product than the one Anthropic launched in 2024. It’s worth a second look.
For the rare reader who can only pay for one AI coding tool: Claude Code is mine. For everyone else, the honest answer is “use both Claude Code and Cursor and pick per task.” They complement each other better than either competes.