Cursor 3 vs Claude Code 2.1, out of the same developer reviews

Cursor 3 and Claude Code 2.1 are both agentic coding tools that read a repository and change files in it, sold in different shapes. Cursor is an editor plus a cloud agent platform priced per person at $20, $60 or $200 a month, with on-demand billing after a plan's included model usage is spent (cursor.com/pricing, read 21 September 2026). Claude Code runs in a terminal, in IDE extensions, in a desktop app and on the web, and is included in Claude's paid plans at $20, $100 or $200 a month, on limits that reset every five hours and carry a weekly cap shared with chat (claude.com/pricing, read 21 September 2026).

Version drift shapes this reading. Cursor's stable track was on 3.16.17 in mid August 2026 and Projects entered beta on 10 September 2026; the Claude Code release we can point at is v2.1.278, published 19 September 2026 (GitHub releases API). For Cursor we read the maker's pricing page, three blog posts and its changelog, one TechCrunch report from 5 March 2026 and three community threads, one on Hacker News (2 April 2026, 544 points, 406 comments). For Claude Code we read its documentation and pricing page, two GitHub API calls on 21 September 2026, three publication pieces (The Register, 9 March 2026; ZDNET, 13 August 2026; InfoWorld, 2 September 2026) and one Hacker News thread (6 April 2026, 1,364 points, 753 comments). No editorial review of either tool at the version named above was readable, so nothing here is a verdict on 3.16.17 or on 2.1.278. Nothing was run or tested.

Consensus

Both tools are described as having moved the developer from writing lines to reviewing output, and both makers now sell agent capacity rather than completion. Cursor states that agent users outnumber Tab users two to one, that more than a third of the pull requests it merges are created by agents on their own cloud machines, and that new Projects users merge 30 per cent more pull requests while Projects-first users merge six times as many (cursor.com/blog/third-era, February 2026, and cursor.com/blog/projects, 10 September 2026, the maker's claims). Anthropic's reported figures for its Code Review service are findings in 84 per cent of pull requests above 1,000 changed lines, with fewer than one per cent of its findings rejected by developers (Anthropic's figures, reported by The Register, 9 March 2026). Both sets come from the companies selling the tools, and we have taken them as claims.

Independently of the makers the agreement is narrower: both tools carry multi-file work that one chat window cannot, and both put a review step in front of the developer. A ZDNET survey of 138 developers that names no version, published 13 August 2026, found three in four choosing Claude Code over its closest rival agent, while a Cursor community thread from 6 April 2026 describes the new agent workspace as not yet usable as a primary editor. Our reading is that the two agree on the destination and disagree about the surface.

Recurring strengths

Cursor's strengths in the sample are one body of work spread across several repositories from one window, cloud agents that keep running when a laptop closes, and the option to drop back into the classic editor (cursor.com/blog/cursor-3, 2 April 2026, the maker's description). TechCrunch reported the same shift on 5 March 2026.

Claude Code's strength is that its terminal shape travels. An InfoWorld column of 2 September 2026 describes one writer shipping an application to Cloudflare Workers inside 48 hours with about four hours of his own work in it. The ZDNET survey of 13 August 2026 reports three in four of its 138 respondents choosing Claude Code, and its key takeaway is that workflow and ecosystem matter more than benchmarks. The repository figures read on 21 September 2026 are 147,101 stars, 24,050 forks and 12,202 open issues, which measure attention, not quality.

Recurring complaints

A Cursor plan includes a set amount of model usage, and what follows it is billed: the maker's pricing page says on-demand usage continues after that amount is consumed, billed in arrears (read 21 September 2026). A Cursor community thread opened on 6 April 2026 lists friction points its author says make Cursor 3 hard to use as a primary IDE, and a bug report on the same forum dated 15 August 2026 describes 3.16.17 crashing at startup on Ubuntu 24.04, with a second poster confirming the same crash on the same version. Two reports describe a problem; they are not a rate.

Claude Code's recurring cost in the sample is its limits. The maker's pricing page states that activity across chat, desktop, mobile and Claude Code draws on one pool, that a rolling five-hour window and a weekly cap both apply, and that reaching a limit means waiting for the reset or turning on credits billed at API rates (read 21 September 2026). The most discussed developer thread we found is titled around Claude Code being unusable for complex engineering tasks after its February 2026 updates; it reached 1,364 points and 753 comments on Hacker News on 6 April 2026, which counts a discussion, not affected developers. Version churn is shared: Claude Code published v2.1.278 on 19 September 2026 (GitHub releases API), and Cursor's changelog carries dated entries for 22 and 28 July, six dates in August and 2 and 10 September 2026 (cursor.com/changelog, read 21 September 2026).

Where reviewers split

The maker against some of its own users. Cursor's announcement presents the workspace as the end of micromanaging agents and jumping between terminals, tools and windows (cursor.com/blog/cursor-3, 2 April 2026), while the community thread opened four days later says that release is blocked by usability gaps. The Hacker News launch thread of 2 April 2026, at 544 points and 406 comments, holds both positions. Our read: a surface some of its first users cannot yet work in is a young surface, not a settled one.

Price against preference. The Register's account of 9 March 2026 leads on cost and speed, reporting Anthropic's figures of 15 to 25 dollars per pull request and about 20 minutes per review. The ZDNET survey of 13 August 2026 leads instead on preference, with three in four of 138 developers choosing Claude Code. The two measure different things, and both were written before the 2.1.278 release.

Craft against throughput. The InfoWorld column of 2 September 2026 has its writer ship an application in a weekend and add that he did not stop to examine the makeup of the code, and that he will miss doing the work. That is one professional's opinion, written in the 2.1 line before the release named above, and we have not counted it as evidence about that release.

Where the sample is thin. None of the three publication pieces reviews 3.16.17 or 2.1.278 as shipped; each appears above with its date and the version it covered. The developer threads carry more of the disagreement than the publications do.

Who each one is for

Cursor suits you if the work happens in an editor with diffs in front of you, if one body of work spans several repositories, and if you want agents that keep going when your laptop closes. It also suits you if a plan with a set amount of included model usage, and on-demand billing after it, matches how you pay for tools (cursor.com/pricing, read 21 September 2026).

Claude Code suits you if your work happens in a shell, over SSH or inside a container, if you want the same agent in a terminal, in an IDE and in CI, and if you already pay for a Claude plan whose five-hour window and weekly cap you have checked against your own use (claude.com/pricing and the Claude Code documentation, both read 21 September 2026).

Neither is declared the winner, because the sample supports a reader's habit rather than a product.

One verdict a week: the most useful Master Review we finished, the complaint that kept appearing, and who should skip it. Get the weekly verdict.

Sources

1. Cursor pricing, official, read 21 September 2026.
2. Cursor changelog, official, entries dated 22 July to 10 September 2026, read 21 September 2026.
3. Meet the new Cursor, official, 2 April 2026.
4. The third era of AI software development, official, February 2026.
5. Introducing Projects, official, 10 September 2026.
6. Cursor is rolling out a new kind of agentic coding tool, news, 5 March 2026.
7. Cursor 3 discussion, forum, 2 April 2026, 544 points and 406 comments.
8. Cursor 3.0 agentic IDE: huge potential, blocked by usability gaps, forum, 6 April 2026, 870 views.
9. Version 3.16.17 crashing on startup in Ubuntu 24.04, forum, 15 August 2026.
10. Claude Code documentation overview, official, read 21 September 2026.
11. Claude plans and pricing, official, read 21 September 2026.
12. Claude Code v2.1.278 release, official, published 19 September 2026, read 21 September 2026.
13. anthropics/claude-code repository figures, official, read 21 September 2026.
14. Anthropic debuts Code Review for teams, enterprises, news, 9 March 2026.
15. 75% of developers I surveyed prefer Claude Code, editorial, 13 August 2026.
16. Claude Code has left a little hole in my soul, editorial, 2 September 2026.
17. Claude Code unusable for complex engineering tasks with February updates, forum, 6 April 2026, 1,364 points and 753 comments.

Method: 17 sources, dated between February and 21 September 2026, read and synthesised with AI on 21 September 2026 across both tools; nothing was run, installed or benchmarked.