Cursor vs Claude Code

Most of the confusion about these two tools comes from one mix-up: people compare Claude Code to Claude, or Cursor to Claude, as though they are the same kind of thing. They are not. And once that is clear, the real comparison gets much shorter.

Claude Code is not Claude

Claude is a model. It reads and writes text. On its own it cannot open your repository, run your tests, or edit a file.

What lets a model do those things is a harness: the layer that hands the model tools, decides what context it sees, enforces permissions, and runs the loop. The industry shorthand is agent = model + harness.

Cursor and Claude Code are both harnesses. Claude is a model that runs inside either of them. That is why the question "Cursor or Claude?" does not quite parse, and why the answer to "can I use Claude in Cursor?" is yes — Cursor currently offers Claude Opus 5, Sonnet 5, Fable 5.1 and most of the 4.x line.

So this is not a comparison of models. Both tools can run the same frontier models. It is a comparison of the machinery around them, which is where the differences turn out to be real.

What changed in 2026

Two things worth knowing before you pick a tool your team will live in.

Cursor is owned by SpaceX. Anysphere, the company behind Cursor, is now part of SpaceXAI. Cursor's own model documentation lists SpaceXAI alongside Anthropic, OpenAI and Google as a model provider, and describes the Grok models it ships as "jointly trained by Cursor and SpaceXAI." Reported deal terms — roughly $60 billion in stock, closing in August — come from press coverage rather than a release from either company, so treat the mechanics as reported and the ownership as fact.

OpenAI is withdrawing its models from Cursor. OpenAI has said it is winding down the contract that supplies its models to Cursor, triggered by the change of control. Press coverage puts the proposed shutoff at 12 November 2026, and stresses that OpenAI calls the date proposed. Cursor has not published its own confirmation, and as of this writing its docs still list the full OpenAI lineup with no deprecation notice. Cursor's CEO has put OpenAI models at about 5% of platform traffic. If you want GPT after the wind-down, the reported fallbacks are your own OpenAI API key or the Codex extension running inside Cursor.

Neither of these should decide anything on its own. Both belong in the file when you are choosing a tool your team will use for the next three years.

The real difference: where the work runs

Both tools run agents in the cloud. This is the part most comparisons get wrong, in both directions, so it is worth being precise.

Claude Code cloud sessions run on Anthropic-managed VMs. You start one from the browser, the mobile app, the desktop app, or claude --cloud in your terminal. The environment arrives with Docker, PostgreSQL, Redis, the usual language runtimes and chromedriver, and you can add a setup script. It clones your repo, installs dependencies, starts services and runs your tests. You can pull a running session down into your terminal with --teleport, branch and history intact.

Cursor cloud agents run in an isolated VM too, but with something Claude Code does not document: a full desktop. Cursor's docs say agents "can use a mouse and keyboard to control the desktop and browser, allowing them to interact with the software they build like a human developer" — they start dev servers, open the app in a browser, click through flows, and verify the change before opening a PR. The agent attaches screenshots and videos to that PR, so you can check the work without pulling the branch.

That is a real advantage and it is the honest version of "Cursor stands out." It is not that Cursor has a cloud VM and Claude Code does not. It is that Cursor's cloud agent ships with a desktop and a browser and treats visual verification as the default output.

Claude Code's answer to browser verification is Claude in Chrome, which drives your own browser on your own machine. It clicks, types, reads the console, and records the run as a GIF. Because it is your browser, it is already logged into everything — which is the better end of the trade for anything behind an auth wall, and the worse end for anything you want to run while your laptop is shut. It stops at login pages and CAPTCHAs by design.

So the split is less "who has the cloud" and more:

• Work you want verified visually, while you are away from your machine, on an app that is reachable without your credentials: Cursor's cloud agents.
• Work that needs your real session, your local services, or your machine's state: Claude Code, locally, with Claude in Chrome.

Claude Code covers the away-from-the-desk case differently. Remote Control connects your phone or browser to a session running on your own machine, so you steer your laptop rather than a VM. And routines run on a schedule, an API call, or a GitHub event — each one a full cloud session that can clone, test and open a PR, not just call an API and report.

Skills travel; plugins are how Claude ships them

Agent Skills started at Anthropic and became an open standard: a folder with a SKILL.md file that tells an agent how to do something specific.

The part worth knowing is that Cursor implements it — and reads .claude/skills/ directly. Cursor also reads CLAUDE.md the same way it reads AGENTS.md. So the skills and instructions a team writes for Claude Code largely work in Cursor with no changes.

There is an asterisk. The portable part of the spec is a small set of fields; Claude Code adds extensions of its own that other tools do not read. A skill that stays inside the standard travels. One that uses the extensions does not, or travels degraded.

Where the two diverge is distribution. Claude Code has plugins: a directory that packages skills together with agents, hooks, MCP servers and settings, published through a marketplace and installed with one command. That pairing is the useful part — a skill that needs an external system can ship with the MCP server that reaches it, so one install delivers the procedure and the tools it depends on rather than half of each. Anthropic curates two official marketplaces, and a marketplace is otherwise just a git repo, so a team can host its own. That layer is the harness's own packaging rather than the open format underneath it, so it is the part that does not carry over.

The practical consequence is still the opposite of what teams assume when they standardize on a tool: the investment is in the skills, and the skills are portable, so picking a harness is a more reversible decision than it looks.

We go through skills, plugins and the rest of the configuration surface in Claude Code plugins and skills, explained.

What it costs

The structural difference matters more than the headline numbers. Cursor gives you a pool for its own models — the Grok family and Composer — and bills third-party models, including Claude, at the model's API price out of a second pool. Claude Code includes Claude models in the subscription.

So the comparison depends entirely on which model you intend to run. A team that wants to run Claude all day is buying it at API rates through Cursor and at subscription rates through Claude Code. A team happy on Cursor's house models is getting the cheaper pool.

So which one

We run both, and keep running both. These tools change shape every few months — a capability that decides the question today is table stakes by spring, as the cloud-VM half of this post shows. Part of what a team pays us for is that we have already used the thing they are asking about.

Which means the useful answer is not a winner. It is a way to decide:

• If the work is visual and you want it verified while you are somewhere else, Cursor's cloud agents are built for exactly that.
• If the work needs your machine — your logged-in browser, your local services, your half-finished branch — Claude Code is built for that.
• If you are going to run Claude models all day, check the pricing structure before the feature list. It is the difference that compounds.

And whichever you pick, put the work into your skills and your written instructions rather than into the tool. Those are the part that transfers.

Sources

Checked September 2026. Where a claim could not be confirmed from a primary source, it is described above as reported rather than stated as fact.

• Cursor cloud agent capabilities — desktop environment, browser control, screenshot and video artifacts
• Cursor models — model lineup and SpaceXAI provider labelling
• Cursor pricing — plans and the two usage pools
• Cursor skills — SKILL.md support and the .claude/skills/ directories
• Claude Code cloud sessions — cloud VMs, --cloud, --teleport, no separate compute charge
• Claude Code cloud environments — installed tools and setup scripts
• Claude in Chrome — local browser control and its limits
• Claude Code routines — scheduled and triggered cloud sessions
• OpenAI on Cursor — the wind-down
• Cursor customers will lose access to OpenAI models in November — the proposed date