Claude Code vs Cursor: Pricing Converges as Security Flaws Mount

Same price, different products

Claude Code and Cursor now cost the same to start using, and that makes them harder to compare, not easier. Both products have settled on a $20 monthly entry tier for individual developers.1114 At the top end they also match: each has a $200 tier for heavy users, Anthropic's Max 20x on one side and Cursor's Ultra on the other.17 The real differences show up in what that money buys, how usage is metered, and how much each tool spends to finish a task.

A September 2026 comparison from security firm NeuralTrust adds a third factor to price and performance. It reports that both vendors shipped critical security fixes this year. On Cursor's side, a zero-click sandbox escape called "DuneSlide" was rated CVSS 9.8. On Claude Code's side, a bypass of the consent step for hooks was rated CVSS 8.7.11 NeuralTrust's conclusion is that neither tool is secure out of the box for enterprise use.11 That is probably the most useful takeaway in this year's crowded field of head-to-head comparisons.

Pricing: matched at the ends, apart in the middle

The gaps are in the middle tiers and in team plans. Cursor offers a Pro+ plan at $60 a month. Anthropic's next step after Pro is Max 5x at $100.14 Cursor also has a free Hobby tier. Claude Code has no free option and requires at least the $20 Pro plan.1415

Published team pricing varies widely, and that inconsistency matters to buyers. NeuralTrust and CatDoes both list Claude's standard team seat at $25 and the premium seat at $125. Cursor charges $40 for a standard seat and $120 for premium.1115 Builder.io says full Claude Code access for teams requires the $125 Premium seat. On that basis it calculates that a ten-person team would pay $1,250 a month for Claude Code versus $400 for Cursor.14 An older Morph comparison put the premium seat at $150.19 Another analysis gave annual-billing figures of $20 standard and $100 premium.13 Anthropic has clearly changed its seat structure several times. Teams should check current rates directly rather than rely on any third-party table.

The two products also meter usage differently. Claude Code uses a rolling five-hour window plus a weekly cap. Cursor uses credit pools and bills overages afterward.15 Builder.io adds that Cursor's $20 plan includes $20 in credits, which it estimates at about 225 Claude Sonnet requests.14 For finance teams, Anthropic's model has firm limits, while Cursor's carries the risk of a surprise bill. Developers may see it the other way around.

The benchmark numbers don't line up

Benchmark coverage disagrees more than anything else in these comparisons. The main reason is that comparing an agent built around one vendor's models with an editor that routes between many models doesn't work cleanly.

Tech Insider's May testing gave Claude Code, running Opus 4.6, 80.8% on SWE-bench Verified, against about 74% for Cursor running Sonnet 4.6.12 An October follow-up from the same outlet reported 97.0% for Claude Code with Claude Opus 5 on the vals.ai leaderboard. It put Cursor's most recent reported score at 51.7% with its in-house Composer model.17 These figures show Anthropic's model progress more than any difference between the tools. The test is the same, but the models underneath keep changing.

Cursor's own benchmark is more informative. On CursorBench 4.0, Claude Opus 5.5 at High effort scored 56.0% at $3.97 per task. Cursor's Composer 2.5 scored 27.7%, ranking 50th of 57, at $0.68 per task.11 That is a striking result for Cursor's flagship model on a benchmark Cursor publishes. CatDoes reports a smaller quality gap. In independent cost-per-task data, Claude Code with Opus 4.7 scored 66 and Cursor with Composer 2.5 scored 62, at $4.10 and $0.07 per task. CatDoes notes that the Claude figure predates Opus 5.5.15 Composer 2.5 also posted 79.8% on SWE-bench Multilingual, close to Opus 4.7's 80.1% on the same run.17

The reasonable reading is that Composer is Cursor's low-cost model for routine work. It is not a challenger to frontier models on hard tasks. CatDoes reports that Composer 2.5 is Cursor's default model and is about eight times cheaper per token than Opus 5.5.15 Claude Code's default is Opus 5.5 on every paid plan.15 The two companies are betting on different things. Cursor bets that most edits don't need the strongest model available. Anthropic bets that they do.

The token-efficiency claim has a problem

One figure appears in almost every comparison: on the same task, one tool used about 33,000 tokens and the other about 188,000. Builder.io, Future Proofing and Morph all say Claude Code is the efficient one, about five and a half times leaner than Cursor. They attribute this to Claude Code reading files as needed and holding them in a large context window, while Cursor reloads context at each step.141319 Totalum cites what appears to be the same test with the numbers reversed, saying Claude Code burns more.16 Most sources point the same way, so Totalum's version looks like a transcription error. Still, a single widely repeated test is weak support for any purchase decision.

Future Proofing adds a fair caveat. Narrow edits to a single file can still favor Cursor.13 The split it describes, Cursor for interactive work and Claude Code for long autonomous jobs, is where most of this coverage ends up.

Cursor's editor keeps turning into an agent manager

Cursor's product changes over the past year explain why it no longer looks like "VS Code with a chat box." Cursor 2.0 added Composer and an interface built around agents rather than files.1 Version 2.4 in January added subagents: parallel workers, each with its own context window, which can spawn further subagents for large refactors.2 Cursor 3.0 shipped on April 2 with an Agents Window for running many agents at once across local, cloud and SSH environments. Version 3.1 added tiled panes.10 Cloud Agents run in isolated virtual machines with their own terminal and browser.2 Self-hosted cloud agents became available to Enterprise customers in March.10 Tech Insider reports that version 3.11 in July added Side Chats and Conversation Search.26

Cursor's corporate status is less clear. NeuralTrust describes Anysphere as "a SpaceX company."11 Tech Insider reported in September that Cursor was reportedly being acquired for about $60 billion.22 Both point to a major ownership change, but the details should be treated with caution. Revenue figures also vary between sources, from more than $2 billion in annual recurring revenue as of February12 to more than $500 million in another account.26

On compliance, NeuralTrust credits Cursor with SOC 2 Type II, ISO 27001, ISO 42001 and AIUC-1, compared with SOC 2 Type 2 and ISO 27001 for Claude Code.11 An earlier comparison said Cursor advertised only SOC 2 as of April.26 If both are accurate, Cursor added certifications quickly this year, probably in response to enterprise buyers.

Windsurf is now Devin Desktop

Any Cursor comparison that still describes Windsurf as a budget alternative is out of date. Cognition renamed Windsurf to Devin Desktop on June 2 through an automatic update that kept users' settings. windsurf.com now redirects to devin.ai.2227 The Cascade agent reached end-of-life on July 1. Its replacement, Devin Local, is the default local agent.25 One review says Cascade was rewritten in Rust and that the new agent is up to 30% more token-efficient.28

The price advantage is also gone. In March, Cognition replaced credits with daily and weekly quotas and raised Pro from $15 to $20, the same as Cursor. Existing subscribers kept their legacy pricing.2122 What Devin Desktop offers instead is its own technology. That includes SWE-series models (SWE-1.6 arrived in May), Codemaps for visualizing code structure, an Agent Command Center that tracks local and cloud agents on a Kanban board, and support for the open Agent Client Protocol.2530

Coverage disagrees about when Cognition bought Windsurf. Some outlets give July 2025, right after Google's $2.4 billion licensing deal took the founders to DeepMind.27 Others say December 202521 or a close in late March 2026.29 The direction is not in dispute. Windsurf was a cheaper Cursor alternative and is now the front end for Cognition's agents.28

What it adds up to

The three main tools are moving toward the same idea: the developer supervises a group of agents rather than writing code line by line. Cursor and Windsurf both released major multi-agent updates in April.23 Claude Code comes from the terminal side but is aimed at the same autonomous, long-running work.12

The practical conclusion is that these tools are increasingly used together rather than as substitutes. Future Proofing estimates about $220 a month per senior engineer to run Cursor at $20 alongside Claude Max 20x.13 The bigger concern is the one NeuralTrust raises. Agents that run shell commands, drive browsers and execute hooks are now producing critical-severity flaws in tools from both vendors.11 As these tools act more autonomously, buyers should weigh sandboxing and permission controls as heavily as benchmark scores.