Claude Code vs Cursor vs GitHub Copilot.

How do Claude Code, Cursor and GitHub Copilot compare?

All three can now read a repository, plan a change, edit several files and run commands. What separates them is the workflow each is built around and how each charges for heavy use. Prices are US dollars unless stated.

Which workflow suits your team?

Start with how your developers like to work, because that decides adoption more than any feature list.

• A terminal agent (Claude Code) suits work you can describe and check: “add tests for the invoicing module”, “upgrade this library and fix what breaks”. You review a finished diff. It also runs in GitHub Actions, GitLab CI/CD and Slack, so it fits scripted, repeatable jobs.
• An AI editor (Cursor) suits developers making many small changes who want to watch each edit. Its rules and team marketplace keep conventions inside the editor.
• An IDE assistant (GitHub Copilot) suits teams that want one tool across mixed editors, including Visual Studio and Xcode, with a cloud agent that works on GitHub. Business and Enterprise add central content exclusion for files Copilot should ignore.

These are not exclusive. GitHub lists Claude Code and Codex as third-party agents inside Copilot Pro, and Claude Code ships VS Code and JetBrains extensions.

Levy Street builds with AI coding agents daily. The first playable version of World of ClaudeCraft, our open-source browser MMO, was built in about 48 hours in June 2026 with the Claude Fable 5 coding model, and SPAWN D, our open-source control plane, runs Claude Code and Codex agents on machines you own. We have not published a benchmark of these three tools, so treat the test below as the way to decide.

How do you test the three on your own repository?

Use a copy of one real repository, synthetic data only, and the same three tasks in each tool. Give each a fortnight with two or three developers.

1. Write tests for untested code. Pick a module with thin coverage. Then break the function deliberately: tests that still pass against broken code are worthless.
2. Fix a closed bug. Take a bug your team already fixed, check out the commit before the fix, and give each tool the original issue text. Compare its change with the human fix.
3. A small cross-file change. Rename a concept used in several files, such as a pricing field, and update tests and docs.
4. Score what matters. Count review comments, rework time, failed CI runs and the cost shown in each tool’s usage dashboard. Our lesson on evaluating AI-generated code has a review checklist.

What do they cost for a six-person team?

Worked example: six developers on annual team plans, at list prices on 20 October 2026, before heavy-use charges.

Illustrative add-on: if each Cursor developer sends 3 million input and 300,000 output tokens a month to Claude Sonnet 5.5 at US$2 and US$10 per million, that adds US$9 a developer, or US$648 a year for six. Claude’s page shows a bare “$” with “Prices include 15% GST”; Cursor and GitHub bill in US dollars, so confirm GST and currency conversion on the first invoice.

How do privacy, training and admin settings differ?

Repositories hold more personal information than people expect: test fixtures, logs and customer exports. Treat whatever an agent reads as data sent to its vendor.

• Claude Code. Under commercial terms (Team, Enterprise, API and cloud platforms), Anthropic does not train on code or prompts. Admins can push managed settings, such as permission rules, to every machine. Sessions are cached in plaintext on the laptop for 30 days by default.
• Cursor. With Privacy Mode on, Cursor does not train on your data and holds zero data retention agreements with model providers. With it off, Cursor may store and train on code, prompts and editor actions. Teams can set privacy team-wide.
• GitHub Copilot. Business and Enterprise data is not used for training. Since 24 April 2026, GitHub uses interaction data from Copilot Free, Pro and Pro+ for training unless users opt out.

Under the Privacy Act 2020, a provider that trains on what it receives is using it for its own purposes, which the Privacy Commissioner treats differently from one acting only for you. A developer on a personal plan with training on is the gap to close. Put company work on business plans and check settings centrally.

Which should you choose?

Choose Claude Code if

• Developers want to delegate whole tasks and review finished changes.
• You already pay for Claude Team, so the seat covers chat and coding.
• You want agents in CI, scheduled jobs or Slack, not only in an editor.

Choose Cursor if

• Your team wants an AI-first editor and steers edits interactively.
• You are comfortable with its 2026 ownership change and Grok-centred plans.
• You will cap third-party model spend from day one.

Choose GitHub Copilot if

• Your code lives on GitHub and developers use several different IDEs.
• You want the lowest predictable seat price and central content exclusion.
• You want IP indemnity and organisation-wide policy in one place.

Where coding agents go wrong

• Tests that agree with the bug. An agent writing both code and tests can encode one misunderstanding twice. Write acceptance cases yourself.
• Too much autonomy. Anthropic says Claude Code now runs in auto mode by default on Pro, Max and Team. Keep deny rules and keep production credentials out of reach.
• Untrusted extensions. MCP servers widen what an agent can touch; see MCP for business systems.
• Surprise bills. Third-party models in Cursor and pooled credits in GitHub can overrun. Set caps before rollout.
• Shifting terms. GitHub’s training change and Cursor’s sale both landed in 2026. Re-read terms at renewal.

For deeper detail see our Claude Code and Cursor guides. If you would rather have the system built for you, that is our custom AI development work.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.