Cursor vs VS Code — is the fork worth $20 a month in 2026?
Cursor is a fork of VS Code, so the editor underneath is the same. What you pay for is the AI layer: a house model at $0.55 per agent task, per-hunk multi-file review and a tuned agent loop. VS Code answers with a free Copilot tier — 2,000 completions a month — and bring-your-own-key chat that runs with no Copilot plan and no GitHub account. The question is whether Cursor's agent beats your own API key plus a free editor.
| Category | Winner | Margin |
|---|---|---|
| Price of the editor · before any AI | BVS Code | VS Code is free forever; Cursor's free Hobby tier limits Agent requests |
| Cheapest working AI setup · what $0 gets you | BVS Code | VS Code runs BYOK chat with no plan and no GitHub account, or 2,000 Copilot completions free |
| Agent experience · multi-file autonomous edits | ACursor | Purpose-built agent loop, inline multi-file diff review and a house model tuned for latency |
| Cost per agent task · measured | ACursor | Composer 2.5 measures $0.55 a task at 6.8 minutes — the cheapest and fastest agent in the neutral index |
| Extensions · marketplace access | BVS Code | Microsoft Marketplace in full; Cursor pulls from Open VSX, where not everything is listed |
| Model choice · who you can run | ·Even | Cursor bundles Claude, Gemini, Grok, GPT and Composer; VS Code reaches the same models plus local ones via BYOK |
| Tab completion · as you type | ACursor | Cursor's house completion model is the feature most switchers name; VS Code needs a Copilot plan for inline suggestions |
| Local models · Ollama and self-hosted | BVS Code | VS Code documents local models through BYOK; Cursor is built around its cloud |
| Telemetry control · turning it off | BVS Code | One documented setting switches VS Code telemetry off entirely |
| Team administration · seats and policy | ACursor | Cursor Teams adds SSO, usage analytics, privacy mode and a rules marketplace; VS Code has no seat layer of its own |
| Update cadence · upstream features | BVS Code | VS Code ships monthly; Cursor rebases its fork onto those releases on its own schedule |
| Best overall | ·Depends | See the decision tree below |
CHOOSE A · CURSOR
If the AI is the point, not an add-on.
- Agent — a purpose-built loop with inline multi-file diff review, not a chat panel bolted to an editor
- Tab — a house completion model that predicts multi-line and multi-location edits
- Cheap tasks — Composer 2.5 measures $0.55 an agent task and 6.8 minutes, the fastest in the neutral index
- Bundled models — Claude, Gemini, Grok and GPT under one $20 subscription, no keys to manage
- Teams — SSO, shared rules, usage analytics and agentic code review on the $40 seat
CHOOSE B · VS CODE
If you want to own the AI bill.
- Free — the editor costs nothing and 2,000 Copilot completions a month cost nothing either
- Your key — BYOK chat works with no Copilot plan and no GitHub account, billed by the model vendor directly
- Local models — documented support for locally hosted models through a BYOK provider such as Ollama
- Extensions — the full Microsoft Marketplace, including the ones Open VSX does not carry
- Telemetry — one setting turns it off, and the source of the OSS build is MIT-licensed
| Aspect | Cursor | VS Code |
|---|---|---|
| The editor itselfBefore any AI | $0 · Hobby The full editor with limited Agent requests and access to Composer. No credit card, no expiry | $0 · Free forever Microsoft's build is free under its own product licence; the underlying Code - OSS repository is MIT B wins |
| Free AI allowanceWhat $0 gets you | Limited Agent requests Cursor does not publish a numeric free-tier allowance; the Hobby tier is described as limited Agent requests plus Composer access | 2,000 completions/mo Copilot Free, on automatic model selection, with limited chat and agent usage. Requires a GitHub sign-in B wins |
| Bring your own keyNo subscription at all | ~ Individual only BYOK works on individual plans. Teams and Enterprise workspaces add a Token Rate of $0.25 per 1M tokens on top of your own API bill | ✓ No plan, no account BYOK chat runs with no Copilot plan and no GitHub sign-in, against Anthropic, OpenAI, Gemini, Azure or a custom endpoint. Inline completions and semantic search still need an account B wins |
| Entry paid tierFirst tier worth having | $20/mo · Pro Includes $20 of Other Models usage plus a separate pool of Cursor's own models, extended Agent limits, cloud agents, MCP, skills and hooks | $10/mo · Copilot Pro Unlimited completions and $15 of monthly AI credits, with a model picker covering Claude Haiku 4.5, Sonnet 4.5/4.6/5 and Gemini 3.1 Pro B wins |
| Heavy-usage tierTop individual plan | $60–$200/mo Pro+ at $60 includes $70 of Other Models usage and 3x Pro's Agent limits; Ultra at $200 includes $400 and 20x | $39–$100/mo Copilot Pro+ at $39 includes $70 of AI credits and premium models; Copilot Max at $100 includes $200 of credits and the full catalogue including Claude Opus 5 B wins |
| Team seatsPer user, per month | $40/user · Teams Standard $40, Premium $120 for 5x the Agent limits. SAML/OIDC SSO, Bugbot code review, shared team context, usage analytics, team-wide privacy mode and an internal rules marketplace A wins | $19/user · via Copilot VS Code has no seat layer of its own — team administration comes from Copilot Business at $19 or Enterprise at $39, governed through your GitHub org |
| Measured cost per agent taskArtificial Analysis, house models | $0.55 / task Cursor CLI on Composer 2.5 Fast, at 6.8 minutes a task — the cheapest and fastest configuration in the index. The standard Composer 2.5 build measures $0.08 at the same score A wins | ~ Not measured No neutral leaderboard scores VS Code's built-in AI as an agent. Running your own key, the cost is whatever the model charges — Claude Opus 5 measured $8.23 a task, GPT-5.6 Luna $0.31 |
| Realistic annual costOne developer, moderate use | $240/yr Cursor Pro at $20 a month, with model usage inside the included $20 allowance for most individual workloads | $0–$120/yr Free if Copilot Free's 2,000 completions and BYOK chat suffice; $120 with Copilot Pro. Any API key spend is billed by the model vendor on top B wins |
| Capability | Cursor | VS Code |
|---|---|---|
| Editor core | VS Code fork | The original |
| Price of the editor | $0 (Hobby) / $20 Pro | $0 forever |
| Source licence | Proprietary | MIT (Code - OSS), Microsoft licence on the branded build |
| Extension marketplace | Open VSX + Anysphere replacements | Microsoft Marketplace |
| Every VS Code extension works | ~ most, not all | ✓ |
| Settings and keybindings import | ✓ one-click on first run | ✓ Settings Sync |
| Inline completion | ✓ house model, multi-line | ✓ with a Copilot plan or free tier |
| Agent mode | ✓ purpose-built loop | ✓ built in, Copilot-powered |
| Multi-file diff review | ✓ inline, per hunk | ✓ chat edits with accept/discard |
| House model | Composer 2.5 / 2.5 Fast | ✗ (routes to other labs) |
| Frontier models included | ✓ Claude, Gemini, Grok, GPT | ✓ via Copilot plan or your own key |
| Bring your own API key | ✓ individual plans | ✓ no plan, no GitHub account |
| Local / self-hosted models | ~ cloud-first | ✓ documented, e.g. Ollama |
| Custom endpoint provider | ~ limited | ✓ |
| MCP support | ✓ plus skills and hooks | ✓ MCP servers in agent mode |
| Terminal agent | ✓ Cursor CLI | ✓ Copilot CLI |
| Cloud / background agents | ✓ on paid plans | ✓ Copilot cloud agent on paid plans |
| Automated code review | ✓ Bugbot, usage-billed | ✓ Copilot code review |
| Telemetry off switch | ~ privacy mode on plans | ✓ telemetry.telemetryLevel: off |
| Team privacy mode | ✓ team-wide | ~ via Copilot org policy |
| SSO | ✓ SAML/OIDC on Teams | ✓ through GitHub org |
| Usage analytics for admins | ✓ Teams | ✓ Copilot Business dashboards |
| Shared team rules | ✓ internal marketplace | ~ repository instruction files |
| Runs on Linux, macOS, Windows | ✓ | ✓ |
| Remote / SSH development | ✓ | ✓ first-party Remote extensions |
| Web version in a browser | ✗ | ✓ vscode.dev |
| Upstream release cadence | ~ rebases the fork on its own schedule | ✓ monthly releases |
§ 04 / BENCHMARKS
Why there's no chart here.
No neutral benchmark covers this pair. VS Code is an editor, not an agent, and no neutral leaderboard scores it. Its AI comes from whichever provider you attach — Copilot, or your own Anthropic, OpenAI, Gemini or local key — so a benchmark of VS Code would really be a benchmark of that model. Artificial Analysis does score Cursor's own CLI and house model (38 on the Coding Agent Index, $0.55 a task), and we quote those figures in the pricing table rather than charting them against an empty column.
A · ANYSPHERE
Cursor
The same editor with the AI rebuilt into it — a house completion model, a tuned agent loop, and the cheapest measured cost per agent task in the neutral index.
Strengths
- Tab completion — the house model predicts multi-line and multi-location edits, and it is the feature most switchers name
- Agent loop — built into the editor with inline per-hunk diff review rather than a chat panel beside it
- Cheap tasks — Composer 2.5 Fast measures $0.55 a task at 6.8 minutes, the fastest configuration Artificial Analysis has measured
- No keys — Claude, Gemini, Grok and GPT arrive bundled in the $20 subscription
- Team layer — SSO, shared rules marketplace, usage analytics, privacy mode and Bugbot review on the $40 seat
Weaknesses
- Extensions come from Open VSX, and not every Microsoft Marketplace extension is listed there
- The same publisher.extension identifier can resolve to different code on Open VSX than on the Microsoft Marketplace
- Its house model scores 38 on the Coding Agent Index against 67 for the best vendor agents
- Teams and Enterprise add a $0.25/1M Token Rate even on bring-your-own-key usage
- As a fork it inherits upstream VS Code releases on its own schedule rather than Microsoft's
Best for
- Developers who want the AI layer to be the product, not a plugin
- High-volume small edits where $0.55 a task beats $8
- Teams that want shared rules and analytics around AI use
- Anyone who does not want to manage API keys
B · MICROSOFT
VS Code
Free, open at the core, and now able to run your own key with no subscription and no account — which changes the arithmetic of paying for an editor at all.
Strengths
- Free — the editor costs nothing and Copilot Free adds 2,000 completions a month
- Your own bill — BYOK chat runs with no Copilot plan and no GitHub sign-in, billed by the model vendor at list rates
- Local models — documented support for locally hosted models through a BYOK provider such as Ollama
- Full marketplace — every VS Code extension, including the ones missing from Open VSX
- Control — telemetry off in one setting, and the Code - OSS source is MIT-licensed
Weaknesses
- Inline completions, semantic search and embedding-backed features still require a GitHub account
- No house model, so agent quality is whatever you attach and pay for
- No seat layer of its own — team administration means buying Copilot Business or Enterprise
- The agent experience is a capable chat-and-edit loop rather than a purpose-built one
- The Microsoft-branded download is proprietary; only the underlying repository is MIT
Best for
- Developers who already hold API keys and want to pay only for tokens
- Anyone who needs a specific extension missing from Open VSX
- Local-model and air-gapped workflows
- Teams that want AI governed through existing GitHub policy
01
Developer who already pays for API keys
You have an Anthropic or OpenAI key, you watch the usage dashboard, and you would rather not add a subscription on top of tokens you already buy.
Reasoning: This is the case VS Code changed. Bring-your-own-key chat now works with no Copilot plan and no GitHub account, so the editor costs nothing and your AI bill is exactly the model vendor's list price with no platform margin. Cursor's $20 buys a better agent loop and a house model, but you are paying it on top of tokens you were buying anyway.
Picked
VS Code
Runner-up: Cursor Pro if the agent loop saves you more than $20 of your own time a month
02
Engineer running dozens of agent tasks a day
You delegate constantly — small refactors, test scaffolding, boilerplate — and you care about how fast the loop closes.
Reasoning: Volume favours the tuned product. Cursor's Composer 2.5 Fast measures $0.55 an agent task at 6.8 minutes, the fastest configuration in the neutral index, and it is included in the $20 plan rather than metered against your card. Running the same volume through a frontier model on your own key in VS Code would cost several dollars a task. The $20 pays for itself inside a week at that rate.
Picked
Cursor
Runner-up: VS Code with a cheap model such as GPT-5.6 Luna, measured at $0.31 a task
03
You depend on one specific extension
Your workflow rests on a marketplace extension — a proprietary debugger, a vendor SDK integration, a licensed language server.
Reasoning: Check Open VSX before you switch. Cursor pulls extensions from Open VSX rather than the Microsoft Marketplace, and its own documentation notes that not every Marketplace extension is listed there, and that the same publisher.extension identifier can point at different code. Anysphere publishes replacements for some popular gaps, but if your one critical extension is missing, no amount of agent quality compensates.
Picked
VS Code
Runner-up: Cursor, if the extension you need turns out to be on Open VSX
04
Working under an offline or air-gapped rule
Code cannot leave the network. Anything that calls a cloud API is out, regardless of vendor promises.
Reasoning: VS Code documents locally hosted models through a BYOK provider, so the editor and a local runtime can stay inside your boundary. Cursor is built around its cloud — the house models, the agent and the indexing all live there — so the requirement decides this before features do.
Picked
VS Code
Runner-up: A self-hosted open-weight model behind an OpenAI-compatible endpoint
05
Team of fifteen deciding on a standard
Mixed experience levels, one budget line, and someone has to explain the number to finance.
Reasoning: Cursor Teams is $40 a seat — $7,200 a year for fifteen — and gives SSO, usage analytics, a shared rules marketplace and agentic review. Copilot Business inside VS Code is $19, or $3,420, governed from the GitHub org you already administer. The honest split is that Cursor is worth the premium if the team genuinely works agent-first; if half of them mostly want completions, the cheaper seat covers them.
Picked
VS Code
Runner-up: Cursor Teams for the agent-heavy half of the team, Copilot for the rest
06
Trying Cursor without committing
You are curious but wary of moving your whole setup into a fork.
Reasoning: The switching cost is genuinely low in both directions: Cursor imports VS Code settings, keybindings and extensions on first run, and your VS Code install stays where it was. Run two weeks in Cursor with tab completion on, then two weeks back in VS Code with the same model on your own key. The question that settles it is not which is better on paper but which one you keep opening.
Picked
Cursor
Runner-up: Staying on VS Code and adding Copilot Pro at $10 to compare like for like
Frequently asked.
Common questions about this comparison, with sources where they matter.
Q · 01 Is Cursor just VS Code with AI? +
Structurally, yes — Cursor is a fork of VS Code, which is why your settings, keybindings and most extensions carry over on first run. What is not the same is the AI layer. Cursor rebuilt completion, chat and the agent loop into the editor rather than attaching them as an extension, ships its own Composer models, and reviews multi-file edits inline per hunk. VS Code's AI comes from whichever provider you attach. The fork is real but shallow; the difference that matters is above it.
Q · 02 Can VS Code do everything Cursor does, for free? +
Most of it, with caveats. VS Code has agent mode, chat, inline chat and MCP support built in, plus a free Copilot tier worth
2,000 completions a month. It also supports bring your own key with no Copilot plan and no GitHub account, including local models. What you do not get free is inline completion quality on par with Cursor's house model — inline suggestions require a GitHub account, and the free allowance is finite. And you do not get Cursor's per-hunk multi-file review or its $0.55-a-task house agent. Free covers a lot more than it did a year ago; it does not cover everything.Q · 03 Do all VS Code extensions work in Cursor? +
No, and Cursor says so itself. Cursor installs extensions from Open VSX rather than the Microsoft Marketplace, and its documentation states that most popular extensions are available there but not every Marketplace extension is listed. It also warns that the same
publisher.extension name can point at a different publisher or different code on Open VSX — worth knowing if you have a supply-chain review process. Anysphere publishes first-party replacements for some widely used extensions that Open VSX lacks. Check your critical extensions before switching, not after.Q · 04 Is Cursor worth $20 a month? +
It depends on how many agent tasks you run. If you delegate work constantly, Cursor's house model measures
$0.55 a task and is included in the subscription, so twenty tasks a week is already better value than paying a frontier model per token in VS Code. If you mostly write code yourself and want occasional help, Copilot Free or Copilot Pro at $10 inside VS Code covers it for less. The break-even is roughly the point where you stop typing code and start reviewing it.Q · 05 Which is better for privacy? +
VS Code gives you more direct control; Cursor gives teams more policy. VS Code's telemetry switches off with a single documented setting (
telemetry.telemetryLevel: off), the underlying Code - OSS source is MIT-licensed, and BYOK means your prompts go straight to a vendor you chose. Cursor's answer is organisational: a team-wide privacy mode and SOC 2 certification on the $40 seat. If your concern is what a vendor collects from you personally, VS Code. If it is proving to an auditor that a team's AI use is governed, Cursor.Q · 06 What does Cursor cost beyond the subscription? +
Model usage above the included allowance, at API rates. Pro includes
$20 of Other Models usage, Pro+ $70 and Ultra $400, with Cursor's own models on a separate pool. Beyond that you continue at the same API rates on pay-as-you-go. One line to watch on team plans: Cursor applies a Token Rate of $0.25 per 1M tokens on Teams and Enterprise workspaces even when the tokens are billed to your own API key. On individual plans there is no such surcharge.Q · 07 Can I run both? +
Yes, and many developers do during evaluation. Installing Cursor does not touch your VS Code installation, profiles or extensions, and Cursor imports your configuration rather than moving it. The practical friction is muscle memory rather than files. If you keep both long term, the common pattern is Cursor for agent-heavy sessions and VS Code for everything else — though at that point you are paying $20 for a tool you open half the time, which is worth noticing.