Comparlys

All comparisons

AI Coding Assistants Compared (2026 Edition)

Ten assistants, eleven criteria, one score sheet. Here is what each one gives a developer in 2026, from the free plan to an agent that takes a task and comes back with a pull request.

Updated October 2026

Every assistant here writes code from a sentence and changes several files at once. In 2026 that is where they start, not what sets them apart. What does: where they work (your editor, the terminal, the cloud or a pull request), whose models they run, what they cost, and whether your code has to leave your machines.

They come in three shapes. GitHub Copilot, Gemini Code Assist, JetBrains AI, Amazon Q Developer and Tabnine began as completion plugins for the editor you already use and grew agents. Cursor and Windsurf are editors of their own, built on VS Code around AI. Claude Code, OpenAI Codex and Cline began as agents: you describe a task, and they read the code, edit files and run commands.

We scored all ten on eleven criteria, from the side of the developer who writes the code. The scores reward assistants that are cheap to start, work where you work, offer models from more than one company and let you keep your code on your own machines.

Quick verdict: pick by how you work

Short on time? Start here. The reasons are in the table and reviews below.

Comparison table: 10 assistants, 11 metrics

Scores run from 1 to 10, and higher is always better for the developer. A lower price, more places to work and more control over models and code score higher. Assistants are ranked by their average.

Assistant Free plan Paid plan from Editor support Inline completions Terminal agent Cloud agent Code review Model choice Local models Self-hosting Open source Average Median
1 GitHub Copilot 710101010101010717 8.410
2 OpenAI Codex 386110101047107 6.97
3 Gemini Code Assist 10868105104167 6.87
4 Amazon Q Developer 78881010104117 6.78
5 Cline 5761101110101010 6.57
6 JetBrains AI 710381101107101 6.27
7 Cursor 783101010610111 6.17
8 Windsurf 7861015610111 5.16
9 Tabnine 3688116101101 5.06
10 Claude Code 1861101064161 4.96

How We Rated the Assistants (Methodology)

We scored what one developer gets from each assistant's public pricing and documentation, not what an enterprise contract adds. Prices are monthly, in US dollars, billed month to month. A feature in preview or beta counts when anyone on the plan can turn it on. Each criterion became a score from 1 to 10 with a written rule, and three pillars sit behind the rules:

We do not score the quality of the code an assistant writes. It follows the model more than the tool, the models change every few months, and most of the tools here offer the same ones.

The 11 Core Metrics Explained

Average and median

Assistant reviews

Ranked by average score. Each review covers what the table cannot: what the assistant is built for, who it suits and where it hurts.

1. GitHub Copilot

Best for: developers who want one assistant in every editor they use.

Copilot is the assistant most developers meet first, and in 2026 it covers more ground than any other tool here. It completes code as you type and suggests your next edit, runs as an agent in VS Code, Visual Studio, JetBrains IDEs, Xcode and Eclipse, works in the terminal as Copilot CLI, and takes an issue on GitHub as a coding agent that opens a pull request. It also reviews pull requests.

Copilot Free gives 2,000 completions and 50 chat or agent requests a month. Pro is $10 a month with an allowance of premium requests, which buy turns with the larger models from Anthropic, OpenAI and Google, and the included models keep working when it runs out. The service runs only in GitHub's cloud. The Copilot extension for VS Code is open source, and it is also the one place where Copilot can use a model on your own computer.

Strong

  • Completions, agent, terminal, cloud agent and reviews in one plan.
  • An agent in five editor families.
  • Models from several companies.
  • $10 a month, with a free plan to start.

Weak

  • Premium requests go fast on the larger models.
  • Runs only through GitHub's cloud.
  • Local models only in VS Code.

Pick it if you want one subscription that covers your editor, your terminal and your pull requests.

Skip it if your code may not leave your own servers.

2. OpenAI Codex

Best for: developers who already pay for ChatGPT and want to hand off whole tasks.

Codex is OpenAI's coding agent, in three forms that share one account: an open-source command-line tool, an extension for VS Code and the editors built on it, and Codex cloud, which works on a task in a sandbox and opens a pull request. On GitHub it reviews pull requests, on every one or when someone mentions it, and JetBrains IDEs offer it in their own AI chat.

Codex has no price of its own: it comes with ChatGPT plans, from Plus at $20 a month, with usage limits that reset every few hours. It runs OpenAI's models and has no completions as you type, so many developers pair it with another tool for that. The command-line tool can also run an open-weight model on your own computer, offline included, which is why it scores high on self-hosting.

Strong

  • Terminal, editor and cloud agent on one plan.
  • Pull request reviews included.
  • An open-source command-line tool.
  • Runs local open-weight models.

Weak

  • No completions as you type.
  • OpenAI's models only.
  • No lasting free plan.

Pick it if you pay for ChatGPT already and want an agent that works while you do something else.

Skip it if you want suggestions as you type.

3. Gemini Code Assist

Best for: developers who want the most use for free.

Gemini Code Assist is Google's assistant for VS Code and JetBrains IDEs, and Gemini CLI is its command-line agent; both share one account and its limits. The free version for individuals is the most generous plan in this table: up to 180,000 completions a month, 240 chat requests a day and 1,000 requests a day in Gemini CLI. On GitHub, Gemini Code Assist reviews pull requests for free.

Higher limits come with a Google AI Pro subscription at $19.99 a month, and teams buy Standard or Enterprise seats. It runs Google's Gemini models only. Gemini CLI is open source, and a company can point it at Vertex AI in its own Google Cloud project. Google's cloud agent, Jules, is a separate product with its own sign-up.

Strong

  • The biggest free plan here.
  • Free pull request reviews on GitHub.
  • An open-source command-line agent.
  • VS Code, JetBrains and the terminal on one account.

Weak

  • Google's models only.
  • The cloud agent is a separate product.
  • No local models.

Pick it if you want an agent and completions without paying, or your team runs on Google Cloud.

Skip it if you want to switch between models from different companies.

4. Amazon Q Developer

Best for: teams that build on AWS.

Amazon Q Developer is AWS's assistant. It completes code and runs an agent in VS Code, JetBrains IDEs, Visual Studio and Eclipse, has a command-line agent, and on GitHub it takes an issue, opens a pull request and reviews pull requests. It also knows AWS: it explains your resources, writes infrastructure code and upgrades old Java projects.

The free tier includes inline suggestions and 50 agent requests a month, and Pro is $19 per user a month. It runs Anthropic's Claude models only, with no local models and no way to run it outside AWS's own service. Its VS Code extension and command-line tool are open source. AWS also makes Kiro, a separate AI editor built around written specs, so two AWS products now cover similar ground.

Strong

  • An agent in four editor families and the terminal.
  • Issues to pull requests, and reviews, on GitHub.
  • Knows AWS services and upgrades old Java code.
  • A free tier with a working agent.

Weak

  • One company's models.
  • No local models or self-hosting.
  • Overlaps with Kiro, AWS's other AI tool.

Pick it if your code runs on AWS and you want help with both.

Skip it if you want to pick your own models.

5. Cline

Best for: developers who want any model, local ones included, in an open-source agent.

Cline is an open-source agent for VS Code and the editors built on it, with a JetBrains plugin and a command-line version. It reads the project, plans, edits files and runs commands, asking before each step unless you allow it, and it shows what every task cost. It is the most open tool here: Apache 2.0 code, any model provider, and local models through Ollama or LM Studio.

Cline has no subscription. The extension is free, and you pay the model provider for what you use, through your own key or a Cline account, so a heavy day on a top model can cost more than a month of a plan elsewhere. It has no completions as you type, no cloud agent and no pull request reviews: it is an agent where you work, and nothing more.

Strong

  • Fully open source.
  • Any model, local ones included.
  • No vendor service between you and the model.
  • Shows the cost of every task.

Weak

  • No completions as you type.
  • Paying per use can add up.
  • No cloud agent or reviews.

Pick it if you want full control of models, costs and where your code goes.

Skip it if you want one flat monthly price.

6. JetBrains AI

Best for: developers who live in IntelliJ IDEA, PyCharm and the other JetBrains IDEs.

JetBrains AI is two tools in one subscription: AI Assistant, which completes code, chats and edits inside the IDE, and Junie, an agent that plans and carries out larger tasks. Both use what the IDE already knows about your project. AI Assistant can also host other companies' agents, such as Claude Agent, in its chat, and Junie can take a task on GitHub and open a pull request.

AI Free gives unlimited local completions and a small monthly cloud quota, and AI Pro is $10 a month and part of the All Products Pack. Models come from several companies, and AI Assistant can run local models, offline included; a company can keep it all on its own servers with JetBrains' enterprise edition. What it lacks is reach: it works only in JetBrains IDEs, with no command-line agent and no reviews of pull requests.

Strong

  • Knows your project as well as the IDE does.
  • $10 a month, with a free tier.
  • Local models and an offline mode.
  • Models from several companies.

Weak

  • JetBrains IDEs only.
  • No command-line agent.
  • No pull request reviews.

Pick it if you work in a JetBrains IDE and want the assistant to know it as well as you do.

Skip it if your team also works in VS Code or the terminal.

7. Cursor

Best for: developers who want an editor built around AI, with the best completions.

Cursor is a fork of VS Code rebuilt around AI, and it keeps VS Code's extensions, themes and shortcuts. Its Tab completion predicts whole edits and the next place you will change, its agent works across the project, several agents can run in parallel, and cloud agents take tasks from the web, Slack or GitHub and open pull requests. Cursor CLI brings the same agent to the terminal.

Pro is $20 a month and includes model use at the providers' prices, with models from OpenAI, Anthropic, Google and Cursor itself; more costs extra once you allow it. Bugbot, its pull request reviewer, is a paid add-on. It ranks seventh for what it is not: you have to switch editors to use it, nothing runs on a local model or your own servers, and none of its code is open.

Strong

  • The strongest Tab completion here.
  • Parallel agents and cloud agents.
  • Models from every major AI company.
  • VS Code extensions keep working.

Weak

  • You have to switch editors.
  • Code review is a paid add-on.
  • No local models or self-hosting.

Pick it if you are happy to move to a new editor for the most complete AI workflow in one place.

Skip it if your team is settled in JetBrains IDEs or must keep code in-house.

8. Windsurf

Best for: developers who want an AI editor for less.

Windsurf is an AI editor built on VS Code, made by the team behind Codeium and owned since 2025 by Cognition, the company behind Devin. Its agent, Cascade, follows what you do in the editor and the terminal, Windsurf Tab suggests completions and next edits, and a JetBrains plugin brings Cascade to IntelliJ IDEA and the other JetBrains IDEs.

The free plan has a small monthly allowance of agent credits, and Pro is $15 a month, the cheapest of the AI editors. It offers models from several companies, including Windsurf's own SWE models. There is no command-line agent, cloud tasks go through Devin, a separate product, and pull request reviews come with team plans. Nothing runs on a local model, and the code is closed.

Strong

  • An agent that follows what you do.
  • Tab completions with next edits.
  • $15 a month.
  • A JetBrains plugin as well as its own editor.

Weak

  • No command-line agent.
  • Cloud tasks need Devin, a separate product.
  • No local models or self-hosting.

Pick it if you want an AI editor like Cursor for less a month.

Skip it if you want an agent in the terminal or in CI.

9. Tabnine

Best for: companies that cannot send code outside their network.

Tabnine is one of the oldest AI coding assistants, and today it sells to companies that need control above everything else. It runs as a service, in your own cloud, on your own servers or fully air-gapped, with models you pick from several companies or host yourself. It works in VS Code, JetBrains IDEs, Visual Studio and Eclipse, with completions, chat, agents and a review agent for pull requests.

That control has a price. Plans start at $39 per user a month, there is only a trial to try it, and the review agent comes with the enterprise plans. It has no command-line agent, no cloud agent and no local model on a developer's own laptop, and its code is closed.

Strong

  • Runs on your own servers, air-gapped included.
  • Models from several companies, or your own.
  • Four editor families.
  • A review agent for pull requests.

Weak

  • From $39 per user a month.
  • No free plan, only a trial.
  • No terminal or cloud agent.

Pick it if your security team decides where code may go.

Skip it if you are one developer looking for a cheap assistant.

10. Claude Code

Best for: developers who hand large tasks to an agent.

Claude Code is Anthropic's coding agent. It started in the terminal and now also runs in VS Code and JetBrains IDEs, in a desktop app and on the web, where tasks run in Anthropic's cloud and come back as pull requests. It reads a whole repository, plans, edits files, runs tests and commands, and can be shaped with subagents, hooks, MCP servers and plugins.

It ranks last on this table because the table scores breadth, not how well an agent works through a hard task. Claude Code has no free plan and needs Claude Pro at $20 a month or more, has no completions as you type, and runs Anthropic's Claude models only, though a company can run them through its own AWS, Google Cloud or Microsoft Azure account. Pull request reviews run through a GitHub Actions workflow you add, and the code is closed.

Strong

  • An agent made for long tasks with many steps.
  • Terminal, editors, desktop, web and GitHub on one plan.
  • Subagents, hooks and plugins to shape it.
  • Runs through your own cloud account.

Weak

  • No free plan.
  • No completions as you type.
  • Anthropic's models only.

Pick it if you mostly hand tasks to an agent and review what it did.

Skip it if you want suggestions as you type or a choice of models.

FAQ

Why do Claude Code and Cursor rank so low?

The table scores what each assistant offers and costs: plans, editors, completions, model choice and control over your code. It does not score how well the agent finishes a hard task, which depends mostly on the model and changes with every release. Claude Code does one thing, the agent, and has no free plan, no completions and one company's models. Cursor asks you to switch editors and keeps everything in its own cloud. If those gaps do not matter to you, the ranking should not either.

Which AI coding assistant is free?

Gemini Code Assist gives the most for nothing: completions, chat and Gemini CLI every day. GitHub Copilot, Cursor, Windsurf, JetBrains AI and Amazon Q Developer have free plans with small monthly allowances. Cline is free software, but you pay the model provider, unless you run a local model. Claude Code, Codex and Tabnine have no lasting free plan.

What is the cheapest paid plan?

GitHub Copilot Pro and JetBrains AI Pro cost $10 a month, and Windsurf Pro is $15. Cursor Pro, Claude Pro for Claude Code, ChatGPT Plus for Codex, Google AI Pro for Gemini and Amazon Q Developer Pro sit between $19 and $20. Tabnine starts at $39 per user.

Do they all use the same AI models?

Often, yes. Copilot, Cursor, Windsurf, JetBrains AI, Tabnine and Cline let you pick models from several companies, so the same model can sit behind several different tools. Four run one company's models: Anthropic's for Claude Code and Amazon Q Developer, OpenAI's for Codex and Google's for Gemini Code Assist.

Can I keep my code on my own machines?

Yes. Cline, Codex's command-line tool and JetBrains AI Assistant can work with a model on your own computer, offline included. For a company, Tabnine runs on its own servers, air-gapped if needed, and JetBrains has an enterprise edition for the same. Claude Code can use Anthropic's models through your own AWS, Google Cloud or Microsoft Azure account, and Gemini CLI through your own Google Cloud project.

Do I need an AI editor, or will an extension do?

An extension will do for most developers. Copilot, Gemini Code Assist, Amazon Q Developer, Tabnine and Cline work inside VS Code and JetBrains IDEs. Cursor and Windsurf are separate editors built on VS Code: moving from VS Code is easy, since your extensions come along, and moving from a JetBrains IDE is not.

Why are Lovable, Bolt and Replit not here?

They build a whole app from a description, in their own environment, often for people who never open the code. That is a different job from an assistant that works in your editor and your repository, so they need a comparison of their own.

The bottom line

No assistant is best for every developer. Where you work decides more than any score: one editor or several, the terminal, the cloud or the pull request.

If you want one subscription that covers all of it, GitHub Copilot leads the table by a wide margin. If you already pay for ChatGPT or Google AI, Codex and Gemini Code Assist come with it. For an editor built around AI, look at Cursor and Windsurf; for the most control over models and code, Cline and Tabnine; for long tasks handed to an agent, Claude Code.

Try two of them on the same real task in your own repository, and keep the one whose changes you merge without rewriting.

Help us improve this comparison

Know a product that belongs in the table, a criterion worth scoring, or a number we got wrong? Tell us. We read every message.