Cursor, Copilot, Claude Code or Cline? AI code editors and coding agents compared, November 2026
By Andrew Miller ·
Most developers no longer need to choose an AI editor. They need to choose an agent, and then decide whether it should live in a new editor, inside the one they already use, or in the terminal. In November 2026 the strongest agent is Claude Code (4.7), the best graphical editor is still Cursor (4.5), and the safest company default is GitHub Copilot (4.4). The rest of this guide explains when each of those is the wrong answer, and what has quietly disappeared from the list since last year.
It has been a year of takeovers. SpaceX now owns Cursor, Cognition renamed Windsurf, Anaconda bought Kilo Code, Harness took Augment’s engine, Tricentis bought Tabnine, and two popular open-source extensions stopped development. A tool you chose in 2025 may not be the same product today, so every fact below carries a date or a source.
The short answer
- You live in the terminal and are happy with one model vendor: Claude Code (4.7), or Codex (4.5) if you already pay for ChatGPT.
- You want agents, autocomplete and code review in one graphical editor: Cursor (4.5), provided you mostly use Grok and Cursor’s own models.
- Your company runs on GitHub and wants one contract: GitHub Copilot (4.4).
- You refuse to tie yourself to one AI lab: OpenCode (4.6) in the terminal, Cline (4.2) in VS Code.
- Your team works in IntelliJ or PyCharm: Kilo Code (4.0) or JetBrains AI (3.9) with your own keys.
- You want to hand over tickets and come back to pull requests: Devin (4.1) or Factory (4.0), with a budget cap set first.
AI-first editors: a new editor built around the agent
These replace VS Code rather than plugging into it. You get the tightest integration, and you also accept that the editor’s owner decides which models you can afford.
Cursor (4.5) is still the editor where agents can do the most. Its Agents Window runs several local and cloud agents side by side, and cloud agents work in their own virtual machines. The change this year is ownership. SpaceX closed its $60 billion all-stock purchase of Anysphere on 14 August 2026, and the $8 Plus plan now covers only Cursor and Grok models. Claude, GPT and Gemini are billed at API list price from the first request. OpenAI has proposed 12 November 2026 as the date it ends Cursor’s direct access to its models, so if your team runs on GPT, that is eleven days away. Our Cursor vs GitHub Copilot comparison covers the trade in detail.
Windsurf no longer exists under that name. OpenAI’s reported $3 billion deal to buy it collapsed in July 2025; Google then hired chief executive Varun Mohan and paid about $2.4 billion to license some of its technology, and Cognition, the maker of Devin, bought the rest days later, according to CNBC. On 2 June 2026 Cognition renamed the editor Devin Desktop. We cover it in our Devin review (4.1): one $20 plan now buys the editor and a cloud agent, but no plan, including the $200 Max tier, publishes a quota.
Zed (4.2) is the choice if you mostly want a fast editor. It is written in Rust rather than built on VS Code, so it stays light on an ordinary laptop. The free plan gives unlimited AI with your own keys or with outside agents such as Claude Agent and Codex, and Pro is $10 a month with hosted usage at API list price plus 10%. It does not run VS Code extensions, its own agent trails Cursor’s, and Business still lacks single sign-on.
Kiro (4.0) is AWS’s editor, and it makes the agent write requirements, a design and a task list before it touches code. Reviewers like that; small fixes feel slow. For AWS shops the choice is largely made: AWS stopped selling new Amazon Q Developer subscriptions on 15 May 2026, and its newest models appear only in Kiro. Pro is $20 for 1,000 credits. Individual plans may use your content for training unless you opt out.
Google Antigravity (4.1) comes with Google AI Pro at $19.99 a month, which makes it excellent value if you already pay for that plan. Gemini 4 Argon needs a paid tier, Google publishes no usage numbers, and forum users report lockouts of several days once the weekly cap is spent. On a personal Google account your interactions are used for training and may be read by reviewers unless you change the setting.
Trae, ByteDance’s VS Code-based editor, is the cheap outsider. We have not reviewed it. Its pricing page lists a free plan with 5,000 completions a month, and Pro at $20, Pro+ at $60 and Ultra at $200 a month, each including the same dollar value of model usage. The page names no models. The reason to hesitate is trust: in July 2025 The Register reported that Trae kept sending data to ByteDance servers after users switched telemetry off. ByteDance said the switch only covered VS Code’s own telemetry. We have found no independent retest since.
Extensions: an agent inside the IDE you already use
If your team has years of settings, plugins and muscle memory in VS Code or a JetBrains IDE, this is the lower-friction route.
GitHub Copilot (4.4) is the safe default for companies on GitHub. You can assign an issue to its cloud agent and get a pull request back, and it offers the widest model menu of any tool here, including GPT-6, Claude, Gemini, Grok and Kimi. The catch is billing. Since 1 June 2026 chat and agents spend AI Credits at each model’s token rate, and community threads report Pro users burning a month’s credits in a couple of long agent prompts. Pro is $10 a month with $15 of credits; Business is $19 a seat.
Cline (4.2) is now the default open-source agent for VS Code, partly because its rivals stopped. It shows every edit and command for approval before it runs, works with almost any provider, and costs nothing beyond your tokens. Those tokens are the risk: long sessions on frontier models can pass a $20 subscription quickly, and you see the total afterwards. Telemetry is on by default, and our repository review (4.4) counts three security advisories in 2026, one of them critical.
Kilo Code (4.0) gives you Cline-style control with a free JetBrains plugin and $15 team seats, plus a gateway to more than 500 models at provider prices with a 5% fee on credit purchases. Anaconda bought it in July 2026. It ships fast, which in our reading of its issue tracker also means regressions and out-of-date docs.
Roo Code and Continue are off the list. Roo Code’s repository was archived in May 2026, and Kilo Code is the nearest heir to its mode-based workflow. Continue’s site says it has joined Cursor, and its repository is marked as no longer maintained. Neither is something to install today.
JetBrains AI (3.9) earns its place because its Junie agent can call the IDE’s own inspections, refactorings and test runners. The included credits are thin, 10 a month on the $10 AI Pro plan at $1 each, but Junie now runs on your own key with no markup, and Air hosts outside agents such as Codex and Copilot without a JetBrains AI plan at all.
Tabnine was the privacy-first autocomplete for enterprises that wanted everything on-premises. We have not reviewed it. Tricentis, a software-testing company, bought Tabnine in July 2026 for its Enterprise Context Engine, which it is folding into its testing platform. Coverage of the deal says nothing about the future of the coding assistant, and as of 1 November tabnine.com’s pricing page redirects to a Tricentis sales form. That is not a product to start a rollout on.
Sourcegraph Cody is gone for individuals. Sourcegraph shut down Cody Free and Pro on 23 July 2025 and pointed users to Amp, which has since become a separate company. Sourcegraph (4.2) itself now sells exact code search and context for other agents, Enterprise only, from a $16,000 annual minimum.
Augment Code (3.3) indexed large codebases better than most, but Harness bought its Context Engine, Cosmos and the Auggie CLI on 8 October 2026, and nobody has said in writing what happens to existing subscriptions, unused top-ups or the code indexes. Model calls also carry a 40% fee over list price. We would not sign or renew until that is answered.
Terminal and autonomous agents: the agent is the product
These run in your shell or in the cloud, and they work with whatever editor you like.
Claude Code (4.7) is the agent the others are measured against. It works in your real repository and shell, the same engine runs in the terminal, IDEs, desktop, web and Slack, and skills, hooks, plugins and subagents let a team shape how it behaves. Opus 5.5 has been the default since 22 September. It comes with the $20 Claude Pro plan, but Anthropic publishes no message counts, and the allowance is shared with Claude chat. On API billing, Anthropic’s own cost guide puts the average at about $13 per developer per active day.
OpenAI Codex (4.5) is the agent you probably already own: ChatGPT Plus at $20 includes the open-source Codex CLI (4.5), the IDE extension, the desktop app and cloud tasks. Its local sandbox keeps network access off and asks before risky commands, a better default than most. OpenAI changes models and limits often; GPT-5.5 left Codex for ChatGPT sign-ins on 14 October. See Claude Code vs OpenAI Codex and Claude Code vs Cursor for the head-to-heads.
OpenCode (4.6) is the strongest open-source answer to both. It is MIT-licensed, works with more than 75 providers including local models, and lets you sign in with a Copilot or ChatGPT account. Its local server has needed three security fixes this year, so keep it updated.
Gemini CLI (3.3) is being retired in practice. Google stopped serving it to free users and AI Pro and Ultra subscribers on 18 June 2026 and points them to Antigravity CLI. Aider (3.6), the original terminal pair-programmer, still makes every edit a reviewable git commit, but it has had no commits since May 2026. Keep it if you have it; do not start with it.
Amp (3.8), from Sourcegraph’s founders, is built for teams: shared threads, Linear issues assigned straight to a cloud machine, and no token fee when you bring your own keys or a ChatGPT or Claude plan. You pay for machine time instead, though its pricing has changed several times this year.
Devin (4.1) and Factory (4.0) are the take-a-ticket, open-a-pull-request agents. Devin bundles that with its editor on one $20 plan; Factory runs OpenAI, Anthropic, Google and open-weight models, and has the security paperwork a bank’s security team will ask for, but no free plan and limits it describes only as multiples of each other. If what you want is an app built and hosted rather than code in your repository, Replit (4.4) is a different category and a better fit.
The comparison table
| Tool | Type | Our rating | Starting price | Your own key | Best for |
|---|---|---|---|---|---|
| Claude Code | Terminal agent | 4.7 | $20/month (Claude Pro) | Anthropic key only | Customisable agent work |
| OpenCode | Terminal agent | 4.6 | Free, plus tokens | Yes | Any model, no lock-in |
| Cursor | AI editor | 4.5 | Free; Plus $8/month | Yes | Agents in a graphical editor |
| OpenAI Codex | Terminal agent | 4.5 | Free tier; Plus $20/month | OpenAI key only | ChatGPT subscribers |
| GitHub Copilot | Extension | 4.4 | Free; Pro $10/month | Not published | Companies on GitHub |
| Zed | AI editor | 4.2 | Free; Pro $10/month | Yes | A fast, light editor |
| Cline | Extension | 4.2 | Free, plus tokens | Yes | Approving every step |
| Devin (ex-Windsurf) | Editor and cloud agent | 4.1 | Free; Pro $20/month | Not published | Editor plus ticket agent |
| Antigravity | AI editor | 4.1 | Free; AI Pro $19.99/month | Not published | Google AI Pro subscribers |
| Kilo Code | Extension | 4.0 | Free; Teams $15/user | Yes | JetBrains teams on a budget |
| Kiro | AI editor | 4.0 | Free; Pro $20/month | Not published | AWS shops, spec-first teams |
| Factory | Autonomous agent | 4.0 | Pro $20/month | Yes, capped | Model-agnostic teams |
| JetBrains AI | Extension | 3.9 | Free; AI Pro $10/month | Yes | IntelliJ and PyCharm users |
| Amp | Terminal and cloud agent | 3.8 | Free; $20/month | Yes | Small teams sharing an agent |
| Augment Code | Extension | 3.3 | $20/month | No | Wait for Harness’s plans |
| Trae | AI editor | Not reviewed | Free; Pro $20/month | Not published | Cost-led side projects |
| Tabnine | Extension | Not reviewed | Sales only, via Tricentis | Not published | Existing customers only |
Qodo, often listed alongside these, is now a pull-request reviewer rather than an editor. Its Pro Team plan costs $30 for 2,500 credits, which Qodo says is about 18 reviews a month, and your own model keys are Enterprise-only.
Which one for which developer
A freelancer on one subscription. Start from the chat plan you already pay for. ChatGPT Plus brings Codex, Claude Pro brings Claude Code, Google AI Pro brings Antigravity. Each is a $20 decision you have already made.
A team of 20 on GitHub. Copilot Business at $19 a seat is $380 a month, with IP indemnity and pooled credits so light users offset heavy ones. Add Claude Code or Codex for the few people who run long agent sessions.
A bank or a public body. Cline, Kilo Code or OpenCode pointed at the Bedrock, Vertex or internal gateway you have already approved keeps code with a provider your security team has signed off. Switch telemetry off on every machine first.
A Java or Kotlin shop. Stay in IntelliJ. Kilo Code’s free plugin or Junie on your own key costs less than moving everyone to a VS Code fork.
A team buried in small tickets. Devin, Factory or Amp will turn Linear or GitHub issues into pull requests, but none of them publishes clear limits, so set a spending cap before you assign the backlog.
The catch: almost every price in this guide is a floor. Plans now meter by tokens, credits or rolling windows, and most vendors publish no hard numbers, so a heavy week costs far more than the sticker. Ownership is the other risk. More than half the names above changed hands, changed names or stopped development in the past 16 months, and an acquirer’s priorities can reprice a plan overnight, as Cursor users found in August. None of these assessments comes from our own testing; they come from vendor documentation, changelogs, press coverage and user reports, linked in each review.
Our take: think of the agent as a contractor and the editor as the office. Hire the contractor first, by model and by who already has your billing details, then let it work wherever your team is comfortable. Most developers who already pay $20 for Claude or ChatGPT have their agent, and need no new editor at all.