Cursor Review: The Agent‑First Code Editor After Its X Subscription Shake‑Up

Cursor has spent 2026 transforming from a smart VS Code fork into something closer to a control room for AI agents. It has also changed owners and, as of this month, changed how it charges. On October 12, SpaceXAI folded Cursor into a new bundle called The X Subscription, alongside Grok and X Premium. For a tool that built its reputation on being model-agnostic, that is a big shift.

We reviewed Cursor as it stands in late October 2026, drawing on the official changelog and credible hands-on testing, to answer the question developers are asking: is it still worth paying for?

What it is

Cursor is an AI-native code editor built on VS Code, made by Anysphere. It pairs fast inline completions with agents that can edit many files, run terminal commands and work in the cloud. SpaceX announced in June that it was exercising its right to buy Anysphere at a $60 billion valuation, placing it under xAI, according to Wikipedia's company history. The New Stack reports the deal closed in mid-August.

The biggest product change came earlier, with Cursor 3 on April 2. It introduced a new agent-first interface while keeping the classic editor available alongside it.

Key features

The Agents Window

Cursor 3's centerpiece is a single workspace for running many agents in parallel, across repositories and environments: locally, in isolated git worktrees, in the cloud or over remote SSH. Agent Tabs let you watch several conversations side by side. Commands such as /worktree isolate changes, while /best-of-n runs the same task across multiple models so you can compare results.

Cloud agents and long-running work

Cloud agents run on isolated virtual machines and keep working after you close your laptop, according to DataCamp. Cursor has steadily stretched this idea. Recent releases tracked by Havoptic include Cursor Projects for multi-month coding efforts (September 10), self-hosted machines for running agents on your own infrastructure (September 2), remote control of local agents from an iPhone (October 6) and shareable interactive canvases from cloud agents (October 16).

Design Mode and the built-in browser

Design Mode lets you select a UI element in the integrated browser and pass its markup, styles and position straight to the agent, which is far more precise than describing a button in words. The browser includes Chrome DevTools so agents can check their own work, though DataCamp notes Design Mode works only in Chrome.

Composer models

Cursor trains its own coding models. Composer 2, released in March, was built by continuing pretraining on the open Kimi K2.5 base model and applying large-scale reinforcement learning, per Cursor's technical report. Composer 2.5 followed in May with improvements for long-running tasks and complex instructions. Cursor reported Composer 2 scoring 61.3 on its in-house CursorBench, up from 44.2 for Composer 1.5, but those results have not been independently audited.

Review and security tooling

Cursor now ships bots beyond the editor. Bugbot handles code review on paid tiers, and a September release added Rollouts and Security Review bots for deployment monitoring and code review, per Havoptic. On October 13 Cursor also launched an app marketplace for Origin, SpaceXAI's code-hosting service, with more than 20 integrations.

Pricing

This is where Cursor changed most. Its pricing page now sells The X Subscription in place of the old Pro and Ultra plans, according to CellCog and DevToolsReview, which re-checked prices on October 18.

The catch is third-party models. Under the new plans, Claude, GPT and Gemini usage is no longer included in the subscription fee and is billed at API rates from the first request via prepaid credits, even on Super and Ultra. The New Stack worked through an example: a developer who paid $20 for Pro and used about $20 of Claude would now pay roughly $50 on Premium. Under the old plans, Pro included $20 of API-priced usage and Ultra $400.

There is good news too. SpaceXAI claims at least double the coding usage with its own models compared with comparable older plans, and Teams now start at $30 per user, down from $40. Existing Pro, Pro+ and Ultra subscribers can keep their plans for now, though a support engineer said non-migrating users will eventually be moved with advance notice. Kingy AI adds that Grok chat, image and coding work share one weekly allowance on paid tiers, so heavy use in one area can eat into another.

Hands-on performance

Independent testers broadly agree that Cursor is excellent at large, multi-file work. In one test by The Vibelog, Composer 2.5 carried out a 23-file JavaScript and TypeScript migration while keeping the overall goal consistent. The same reviewer praised in-IDE pull request review, though it could not trigger CI re-runs.

The weaknesses are just as consistent:

• Supervising parallel agents is hard. In The Vibelog's testing, two agents produced code that was internally consistent but incompatible with each other. More agents means more review.
• Cloud is slower than local. Handoffs between environments work, but with noticeable delay.
• Rough edges remain. DataCamp reported environment setup failures, heavy RAM use in long parallel sessions and user reports of unexpectedly high token consumption tied to caching issues.
• Large monorepos still strain it. Legacy codebases can exceed what the context window handles gracefully.

Cursor's original strength still holds up. Its tab completions remain quick and project-aware, and developers who never open the Agents Window can treat it as a familiar VS Code-style editor with very good autocomplete. The learning curve only gets steep once you start orchestrating several agents at the same time.

There are compliance gaps for regulated teams as well. DataCamp notes Cursor lacks FedRAMP and HIPAA certification, and that Workspace Trust is off by default.

Who it's for

• Developers happy with Grok and Composer get strong value, especially on Plus or Premium.
• Teams running large refactors or greenfield builds will get the most from parallel and cloud agents.
• Claude- or GPT-first developers should model their monthly API spend before switching from a legacy plan.
• Regulated industries should verify compliance requirements first.

Alternatives

GitHub Copilot remains the safe choice for teams already living in GitHub, and its agent mode may be enough for many. OpenAI Codex now runs in the cloud and the CLI, with new code review and security scanning announced at DevDay, per InfoQ. Claude Code is a strong terminal-first option for developers who want Anthropic's models without an API surcharge layered on an editor subscription.

Verdict

As a piece of software, Cursor has never been better. The Agents Window, cloud agents and Design Mode make it the most ambitious AI editor around, and the steady release cadence shows no sign of slowing. But the X Subscription turns it into a Grok-first product with a toll on rival models, which undermines a core reason many developers chose it. If you are comfortable inside SpaceXAI's ecosystem, Cursor is excellent value. If Claude or GPT is your daily driver, run the numbers carefully before you migrate.