Cursor is the AI code editor to beat. Where Copilot still wins.

Cursor is the AI code editor that has pulled clear of the field. The San Francisco company behind it, Anysphere, says more than 1 million developers open the app each day, and recent reporting puts annual recurring revenue near $US2 billion. SpaceX reportedly holds an option to buy the firm for $US60 billion. For Australian developers weighing $A30 a month for Cursor against the cheaper GitHub Copilot, the question is no longer whether AI editing is useful, but which tool fits the work. This guide covers what Cursor does, how it stacks up against Copilot, Claude Code and Windsurf, what it costs in Australia, and the trade-offs that come with the choice.

What is Cursor and what does it actually do?

Cursor is an AI-native code editor built as a fork of Microsoft’s Visual Studio Code. Anysphere, the company behind it, was founded in 2023 by four MIT graduates including chief executive Michael Truell and co-founder Sualeh Asif. Functionally it is VS Code with deep model integration. Existing extensions, themes and settings transfer across, but the editing surface assumes a model is in the loop.

Three features carry the work. Tab autocomplete predicts the next several lines from cursor position, file context and recent edits, and accepts on a single keystroke. Agent Mode runs autonomous multi-file changes, opening a sidebar where Cursor proposes an edit plan, applies it across the repository, and stops at user-set checkpoints. Codebase indexing parses the entire workspace into a vector index so the model can reach for relevant files when answering. Users describe the combination as the difference between an autocomplete and a collaborator.

The May 2026 release added PR Review for inline commentary on pull requests, Build Plan in Parallel for splitting an agent task into concurrent threads, and Split PRs to break a long agent change into reviewable commits. Anysphere also ships Composer 2, an in-house model fine-tuned on the editor’s interaction logs. It now handles a growing share of Tab and Agent calls rather than routing every request to Anthropic or OpenAI.

The customer roster covers public references most enterprise software does not get. NVIDIA’s Jensen Huang has called Cursor his “favourite enterprise AI service” and said all NVIDIA engineers use AI coding assistance. Stripe, OpenAI president Greg Brockman and Eureka Labs founder Andrej Karpathy have each surfaced as users.

How much does Cursor cost in Australia?

Cursor bills in US dollars. The free Hobby tier covers limited Tab usage, two weeks of Pro features and a small monthly request quota. Pro at $US20 a month is the headline tier most individual developers use, with extended context windows and unlimited completions. Pro+ at $US60 a month and Ultra at $US200 a month bump the credit pool three-fold and twenty-fold for users running heavy agent workloads. Teams at $US40 per user per month adds single sign-on, admin controls and centralised billing for organisations of three or more.

At the May 2026 exchange rate, Pro lands close to $A30 a month, Pro+ near $A92, Ultra around $A305, and a five-seat Teams plan at roughly $A305 a month all-in. There is no Australian dollar billing option, and no GST receipt is issued automatically. That matters for businesses claiming input tax credits.

For comparison, GitHub Copilot is $US10 a month for individuals, $US19 per user for Business and $US39 per user for Enterprise. Windsurf Pro is $US15 a month with a more generous free tier. Claude Code ships as part of Anthropic’s $US20 Pro and $US100 Max bundles with no separate seat charge.

What is Cursor like to use day to day?

Editing in Cursor feels closer to pair-programming than to typing. A blank file with a single comment describing the desired function will, on a good day, draw a complete first cut of the code on the next keystroke. The experience is less reliable on legacy codebases with idiosyncratic patterns or weak typing, where the suggestions over-confidently produce code that compiles but does not match house style.

Tab autocomplete is the feature most users notice first. It is faster than Copilot’s equivalent and reads more of the surrounding file, which means it is more likely to predict an entire function rather than a single line. The cost is that wrong suggestions can be wrong in larger blocks, so accepting blindly is risky. Most Australian developers describe a posture of skim-and-accept on familiar code and review-line-by-line on anything new.

Agent Mode is where the editor justifies the premium price. Asked to refactor a class across five files, Cursor will open them in a working set, propose the edits in a diff view and apply them on confirmation. Asked to add a feature, it can write the code, generate tests and run them, all inside the IDE. The agent is bounded by configurable rules in .cursorrules files and by user prompts. An agent that has been told not to delete files or run network calls usually obeys.

Codebase indexing is silent until it is needed. Asking the chat sidebar a question about how a particular utility is used will have it cite the calling files by path, jump to definitions, and propose changes that respect existing conventions. On large repositories the index takes minutes to build the first time. Subsequent invocations are quick.

The first wall most developers hit is cost. Pro’s request pool runs out faster than expected on heavy agent days, and the upgrade prompts to Pro+ and Ultra are persistent. The second wall is privacy. Cursor sends source code to its own servers and to whatever model the user has selected. There is a Privacy Mode that disables training-data collection, but it is not the default. Teams in regulated Australian sectors, particularly financial services and healthcare, have to read the data-processing terms carefully before approving the tool.

How does Cursor compare to Copilot, Claude Code and Windsurf?

The four tools are not interchangeable. Each occupies a different point on the autonomy and integration axes.

GitHub Copilot is the safest pick for developers already inside the GitHub ecosystem. It sits in VS Code, Visual Studio, JetBrains and Neovim, has the deepest enterprise compliance story (SOC 2, ISO 27001, GitHub’s own data-handling guarantees), and integrates with pull request review and issue suggestions in a way no third party can match. The autocomplete is slightly less aggressive than Cursor’s, which some developers prefer. Agent Mode shipped to Copilot in 2025 and is now competitive on simple tasks, though it remains a step behind Cursor on multi-file refactors.

Claude Code is the terminal-native option. It runs as a CLI inside any directory, drives the Anthropic Claude models without an IDE wrapper, and reads as the more deliberate hand on architecture and design tasks. It is the AU developer pick for one-off scripts, code review and complex multi-step problems where the developer wants to inspect every file edit at the shell. Without an editor surface, Tab-style autocomplete is not part of the proposition.

Windsurf, owned by Codeium, is the closest direct competitor to Cursor. Like Cursor it is a VS Code fork. Unlike Cursor it ships an agentic-default experience and a meaningfully more generous free tier. Windsurf’s appeal in Australia is the on-premise option for enterprises that cannot send code to the public cloud. The flip side is a smaller model menu and a less polished agent experience on long-running tasks.

ChatGPT, Gemini and the open-weight models such as DeepSeek all show up in AU developer workflows, but rarely as the primary editor. They are the place developers ask broader questions, generate one-off scripts or experiment with architecture, then paste the result into the editor of choice.

The pragmatic Australian setup most experienced developers settle on is some mix of Cursor or Copilot for in-IDE editing, Claude Code in a terminal for design conversations and one-off tasks, and a chat tool such as ChatGPT for the work that does not fit either.

What are the trade-offs?

Cost is the most visible. A heavy user on Pro will exhaust monthly credits in two to three weeks of agent work and either upgrade or wait for the next billing cycle. The Ultra tier at $US200 a month is the only one that comfortably handles full-time agent use, and even that has limits.

Vendor lock-in is the second concern. Because Cursor is a VS Code fork rather than a plugin, every developer on a team has to install Cursor, switch over their settings, and keep up with Anysphere’s release cadence. Going back to vanilla VS Code is straightforward, but the muscle memory of Tab-accept and the agent sidebar takes time to retrain.

The privacy posture is acceptable but not industry-leading. Source code is transmitted to Anysphere infrastructure for processing and to the underlying model providers (Anthropic, OpenAI, others) for completion. Privacy Mode, available on paid plans, prevents Cursor from retaining or training on the data, but the request still leaves the developer’s machine. Self-hosted is not on the roadmap as of mid-2026, so Australian organisations with a strict data-locality requirement will not be able to use the cloud-only product.

Output quality varies by language and codebase age. JavaScript, TypeScript, Python and Go are well-supported. Less common languages get less reliable results. Older codebases with unusual conventions, generated boilerplate or weak test coverage tend to draw plausibly-looking but wrong suggestions, which a less experienced developer can miss.

Finally, there is dependency risk. Anysphere is a young company, well-funded but yet to demonstrate a sustainable margin profile at scale. The reported SpaceX acquisition option introduces a different kind of risk. An editor under Elon Musk’s umbrella may not be a comfortable choice for organisations that have already drawn lines around X and Tesla.

Should you use Cursor?

For Australian developers who write code as their day job, the answer in May 2026 is yes, with caveats. Cursor is the most productive editor on the market for greenfield work, multi-file refactors, and the kind of medium-complexity feature work that fills most engineering backlogs. The $US20 Pro tier pays for itself in a few hours of saved typing for almost any individual contributor.

Three groups should pause before adopting. Solo developers on hobby projects do fine with the free tier of Copilot or Windsurf and do not need the firepower. Engineers at organisations with strict data-residency or self-hosting requirements should look at on-premise Windsurf or stay with Copilot Enterprise. Architects, technical writers and code reviewers who spend most of their day reading rather than editing will get more value from Claude Code at the terminal than from any IDE-bound tool.

Teams adopting Cursor should pair it with explicit .cursorrules files, a documented Privacy Mode policy, and a budget allocated to Pro+ or Ultra seats for the engineers who use Agent Mode heavily. Without those guardrails the tool either gets disabled by IT or burns through credits unevenly across the team. AU enterprise software is absorbing agentic AI fast. Atlassian’s Teamwork Graph being opened to third-party agents at Team '26 was a recent example.

Frequently asked questions

Is Cursor better than GitHub Copilot?

For multi-file refactors, agentic work and codebase-aware questions, Cursor is meaningfully ahead. For simple autocomplete in a familiar VS Code or JetBrains setup, the gap is narrower, and Copilot’s enterprise tooling tilts the balance back. Most Australian developers who switch describe the gap as a step change, not a marginal upgrade, on the kind of work where both tools compete.

Can I use Cursor offline or self-hosted?

No. Cursor runs the editor locally but routes all model calls to cloud servers (Anysphere’s own and the upstream model providers’). There is no on-premise or air-gapped option. Australian organisations with strict data-residency requirements should look at Windsurf’s enterprise on-premise tier or stay with Copilot Enterprise’s compliance story.

Does Cursor train on my code?

By default, Anysphere may use anonymised request data to improve Cursor’s models. Privacy Mode, available on paid plans, disables that retention and training. The default-on data-collection posture has drawn criticism, and teams handling sensitive data should turn Privacy Mode on at the organisation level rather than rely on individual developers.

What models does Cursor use?

A user-selectable mix. Anysphere’s in-house Composer 2 handles a growing share of Tab and Agent requests. For chat and complex Agent runs the menu includes Claude Sonnet, Claude Opus, GPT-5 and others. Switching models is a dropdown in the chat sidebar.

How does Cursor compare to Claude Code?

Different shapes. Cursor is an editor with deep autocomplete and an agent sidebar. Claude Code is a CLI that operates on whatever directory the developer is in. Many Australian developers run both. Cursor handles IDE-bound editing while Claude Code covers design conversations, code review and one-off scripts at the terminal.