Skip to content

Osum · Updated

Claude Code vs Cursor vs Gemini CLI (2026)

Every coding agent demos well. The second week decides. Claude Code, Cursor and Gemini CLI all read your repo and write real code, and the differences that matter are where each one runs, what your subscription already covers and how much of the work you watch as it happens. Here is how the three compare on setup, cost, control and the jobs each one handles well, plus what Claude Code really costs in 2026.

The short answer

Pick Claude Code if you want an agent that edits across files and runs your tests in a loop, from the terminal, an IDE, the desktop app or the web, and you already pay for Claude Pro or Max. Pick Cursor if you want every change inside an AI-first editor, to accept or reject each edit as it appears. Pick Gemini CLI if you want a capable terminal agent on a free tier and an open-source codebase you can read.

All three demo well on a fresh repo. Week two decides it. The people who ship with any of them cut the work into small slices, give the agent a check it can run and read every diff before it goes anywhere. The ones who hand over a whole spec write posts like this one: “Literally nothing working correctly lol. It was rough.”

The short answer
JobWhat users reportYour decision
A bug in an existing repoOne engineer isolated two infrastructure bugs in 90 minutes, a job they’d estimated at days by handStart here. Reproduce it with a test first, then let it fix.
A whole app from one specA thorough spec handed over in one go, deployed, and nothing workedDon’t. Build one slice, test it, then the next.
An app you started in Claude chatOne user rebuilt theirs in Claude Code on 11% of their weekly usageMove the project into Claude Code instead of pasting code into chat.
Non-code work: documents, data, file cleanupSome non-developers now use it for documents tooFine, on a copy of the files.

What’s new in Claude Code for 2026

Claude Code now runs on five surfaces: the terminal, VS Code and JetBrains, a desktop app with a Code tab, and the web at claude.ai/code, including the Claude mobile apps. Your CLAUDE.md, settings and MCP servers carry across all of them. You can start a task on the web and pull it into your terminal, or schedule recurring work as routines that run in the cloud.

The biggest change for new users is the starting permission mode. From version 2.1.283, interactive terminal and VS Code sessions start in auto mode, where a second model, a classifier, reviews actions instead of you. That’s fast, and it means Claude Code can edit and run commands without asking. Press Shift+Tab to see your mode and switch to plan or Manual when the work is sensitive.

Anthropic’s pricing page also now says “Claude Cowork is now just Claude,” rolling out to Pro and Max. The coding product is still Claude Code.

Claude Code vs Claude chat vs Cowork

One user put it simply: “The model is the same. The workflow is completely different.” In Claude chat you paste code in and copy answers out, and every message carries everything you pasted. Claude Code works on your files directly. It reads the files a task needs, edits them, runs your tests and reads the result, without you shuttling code back and forth.

Cowork is the same agent idea in the Claude app, pointed at documents and folders rather than code, with no terminal. If the word “terminal” is what’s stopping you, you don’t need one for Claude Code either: the desktop app includes it in the Code tab.

A simple rule from one owner who uses both: “if I'm still figuring out what I want, I'm in Claude. Once I know what I want, Claude Code gets it done.”

Claude Code vs Cursor, Gemini CLI and the rest

All of these read a codebase, edit several files and run commands. The real differences are where they run, how you pay and how much you see while they work. Claude Code and Gemini CLI are agents you steer from a terminal and several other surfaces. Cursor sits inside its own editor first. Codex and Copilot bring an agent to ChatGPT and GitHub.

Prices below were checked on 28 September 2026 on each provider’s own page. They change often, so check the linked page before you buy. If you want the other Claude-vs-Codex pairing, the Codex comparison owns it: it opens from that course’s page below.

Claude Code vs Cursor, Gemini CLI and the rest
OptionWhere it runsPaid fromBest atWatch out for
Claude CodeTerminal, VS Code, JetBrains, desktop app, web and mobileClaude Pro, $20 a month or $17 billed yearly (checked 28 September 2026)Multi-file changes, bug hunts, work that runs tests in a loopShares usage limits with your other Claude use. Recent versions start in auto mode.
CursorIts own AI code editor, plus cloud agents and a CLIPro, $20 a month; Hobby is free with limited agent requests (checked 28 September 2026)Watching and accepting each change inline as you codeEach plan includes set usage; on-demand usage after that is billed in arrears.
Gemini CLITerminal, with a GitHub Action for pull request reviewsFree with a personal Google account: 60 requests a minute, 1,000 a day (checked 28 September 2026)A free, open-source terminal agent with a 1M token context windowFree-tier quotas can run out on a heavy day; paid use goes through Gemini API or Vertex AI.
OpenAI CodexCLI, IDE extension, desktop app, web and iOSIncluded in ChatGPT. Plus is $20 a month; Free and Go ($8) cover lighter use (checked 28 September 2026)Cloud tasks in parallel, automatic code review on GitHubUsage limits are shared with other agent features in ChatGPT.
GitHub CopilotVS Code, JetBrains, GitHub and a CLIFree tier. Pro $10 a month, Pro+ $39 (checked 28 September 2026)Autocomplete and agent tasks inside GitHubUsage draws on monthly AI credits.

Claude Code vs Cursor: where you watch the work

Cursor is better when you want to stay inside one editor and judge every change as it appears. Its inline diff flow is the product: the agent proposes edits beside your code, you accept or reject line by line, and Hobby gives you a free way to try that loop. Developers who review as they go feel at home there.

Claude Code is better when the job is bigger than one editor session: a bug that spans five files, a refactor with a test to rerun, a change you want planned first and built while you step away. It runs your tests in a loop until they pass, works across terminal, IDE, desktop, web and mobile, and starts from a plan you approve in plan mode.

Choose Cursor if you want every edit to pass in front of your eyes inside its editor. Choose Claude Code if you want to hand over one tested slice and review the diff at the end.

Claude Code vs Gemini CLI: what you already pay for

Gemini CLI is Google’s open-source terminal agent, Apache 2.0 licensed, and its free tier is the headline: sign in with a personal Google account and you get 60 requests a minute and 1,000 a day, with Gemini 3 models and a 1M token context window. You can read every line of its code, extend it with MCP servers and run it in scripts. For a developer who wants a capable agent at no cost, it is a real option.

Claude Code is the paid, managed counterpart: no quota math across your day, deep editor and mobile surfaces around the terminal, plan mode, permission rules with deny enforcement, CLAUDE.md project memory, skills, hooks and checkpointing. You pay with a Claude plan instead of watching daily request counts.

Choose Gemini CLI if your jobs fit the free tier and you value open source over hand-holding. Choose Claude Code if you want the guardrails, the surfaces and the usage your plan already includes, and you would rather set deny rules once than ration requests.

What Claude Code costs right now

Claude Code isn’t in the free Claude plan. It comes with Pro, Max and Team seats, or you can pay per token through an Anthropic Console account. The prices below were checked on 28 September 2026.

Every plan has usage limits, and Claude Code draws on the same allowance as the rest of your Claude use. Anthropic’s docs say it plainly for cloud sessions: they “share rate limits with all other Claude and Claude Code usage within your account.” Run /usage to see your plan’s usage bars.

Three habits stretch any plan. Run /clear between unrelated tasks, because stale context costs tokens on every later message. Use Sonnet for most coding and save Opus for hard reasoning; Anthropic’s own guide says Sonnet handles most coding tasks well and costs less. And point Claude Code at the files a task needs instead of pasting your whole app into chat.

What Claude Code costs right now
PlanPriceClaude Code
Free$0Not included
Pro$20 a month, or $17 a month billed yearly ($200 up front)Included
Max 5x$100 a monthIncluded, 5x Pro usage
Max 20x$200 a monthIncluded, 20x Pro usage
Team, standard seat$25 per seat a month, or $20 billed yearlyIncluded
Team, premium seat$125 per seat a month, or $100 billed yearlyIncluded, 5x standard seat usage

Where Claude Code breaks after the first days

The spec handoff. A data engineer who says they can’t live without Claude Code teamed up with a product manager, handed it a thorough spec and told it to implement everything and run the tests. Then: “Deployed to Railway. Went to try it. Literally nothing working correctly lol. It was rough.” The top reply blamed two words, “implement everything”, for baking in too many assumptions. A whole spec is mostly gaps, and every gap becomes a guess.

The usage burn. People who build in chat paste their whole app into every message. One wrote: “Burned my entire weeks usage in about 5 hours.” After switching to Claude Code with Sonnet, the same person wrote: “I’ve used 11% of my weekly usage and my app is done!” Claude Code reads the files a task needs. Long sessions still carry dead weight, which is why /clear and /compact matter.

The forgetting. Each session starts fresh. Without notes, it reopens decisions you already made. One builder’s fix: “writing really detailed CLAUDE.md specs with architecture decisions and constraints, then breaking the PRD into tiny vertical slices”.

The patch spiral. A veteran engineer noticed it “tends to fix by incremental patching” instead of stepping back. Anthropic’s own advice fits: if you’ve corrected it more than twice on the same issue, clear the session and start again with a sharper prompt.

The Claude Code brief template: slice, check, finish line

The fix you can run today is the brief from lesson 4 of the course, the one that turns a big build into one slice Claude Code can finish and you can check. Paste it with your own brackets filled:

“Plan one slice: [the smallest version a user could use]. It touches [files] and must not touch [out of scope]. Finish line: done when [test] passes and nothing else changed. Explore in plan mode first, and save the plan to plans/slice-1.md.”

Worked example: a founder with a four-page booking-app spec asked for the smallest slice a user could click through: pick a time slot and see it confirmed. Three files, one new test, no new dependencies, and a proposed database migration cut from the slice. Claude Code built one working flow instead of half an app.

  1. The plan names real files, steps and a test.
  2. The slice is small enough to finish in one session.
  3. The finish line says what must pass and what must stay unchanged.

The one-slice workflow people who ship use

The pattern that keeps coming up from people who ship with Claude Code is the same: “repeatedly test and fix. Like, continuously.” Here it is as steps you can run tonight.

  1. Explore in plan mode. Press Shift+Tab until you see plan mode, then ask it to map the files, callers and tests for the change.
  2. Plan one slice. Something a user could click through, with the files it will touch and the test it will run. Save the plan to a file.
  3. Write the failing test first. Watch it fail for the right reason before any fix.
  4. Let it build and rerun until the same test passes, then run the full suite.
  5. Review the diff in a fresh session. Read git status too, for new files the diff doesn’t show.
  6. Commit locally, run /clear, and start the next slice from the saved plan.

Claude Code FAQ

How much does Claude Code cost? It comes with Claude Pro at $20 a month, or $17 a month billed yearly. Max is $100 or $200 a month for 5x or 20x Pro usage. It isn’t in the free plan.

How do I install Claude Code? On macOS, Linux or WSL run curl -fsSL https://claude.ai/install.sh | bash, or use brew install --cask claude-code. On Windows use the PowerShell installer or winget install Anthropic.ClaudeCode. Then type claude in your project folder and sign in.

Can I use Claude Code without the terminal? Yes. The Claude desktop app includes Claude Code in its Code tab, and you can also use it in VS Code, JetBrains or on the web.

Does Claude Code ask before it changes my files? It depends on the permission mode. Manual asks before edits and commands, plan mode explores without editing, and auto lets a classifier approve actions for you. Recent versions start in auto, so check with Shift+Tab. Deny rules block an action in every mode.

Can I undo what Claude Code did? Press Esc twice with an empty prompt, or run /rewind, to restore code, conversation or both to an earlier point. Checkpoints track edits made by Claude’s file-editing tools, not every shell command or outside effect, so keep working on a git branch too.

Does Claude Code use less usage than Claude chat? It can. In chat you resend your code with every message. Claude Code reads the files a task needs. Long sessions still add up, so clear between tasks.

Should I use Claude Code or Cowork if I don’t code? For documents and folders, Cowork is simpler. Some non-developers prefer Claude Code for the same work once they’re comfortable with it.

Verdict: which AI coding agent should you use?

Claude Code is the best pick for most people who want an agent working on their real repo this week: it runs wherever you work, it reads and edits across files, and it runs your tests in a loop. Choose Cursor if you’d rather judge every change inline in an editor built around that loop. Choose Gemini CLI if a free, open-source terminal agent fits your day and your jobs fit its free tier.

Whichever you pick, the skill that decides whether it saves you time is the same: briefing the agent with the right context, cutting work into slices it can finish, giving it a check it can run and reading the diff before you ship. We call it the Checked Slice. If you’re going with Claude Code, our course teaches it in 12 hands-on lessons on your own project, with a certificate and every other Osum course included.

Keep babysitting Claude Code. Or hand it one tested slice and get code you’d ship.

The Checked Slice is four habits: brief one slice, hand it a check it can run, read the diff, send back one line. You get a code map in lesson 1, a reviewed commit by lesson 12, a certificate and every other Osum course. A few quick questions, then your plan.

Get my Claude Code plan About 2 minutes. Lesson 1 is a code map you can check yourself the same day.

Keep exploring

Related learning paths

Learn a useful skill. Put it to work.

Practical courses for AI tools and skills. Course and tool names are the trademarks of their respective owners; Osum is independent and not endorsed by them.