The best AI coding agent in 2026: 9 compared by score, price and limits

Disclosure: Enjoy is our product. It isn’t a coding agent: it runs six of the agents below. We checked every price and score against each maker’s own pages and Artificial Analysis’s index on November 1, 2026, and we say where you don’t need Enjoy at all, which is most of the time.

The best AI coding agents, compared

A coding agent reads your project, edits files and runs commands until a task is done. No agent wins on everything, and nearly all of them come with their maker’s own plans, so choosing one mostly means choosing which plan to pay for. Here are nine side by side, each at its best score on Artificial Analysis’s Coding Agent Index, the nearest thing to a neutral measure.

Agent	Best for	Price, checked November 1, 2026	Index, best setup	Runs on
Claude Code	The hardest work	Claude Pro $20 a month, or $17 billed yearly; Max $100; Ultra $200	61.7, with Opus 5.5	Mac, Windows, Linux; VS Code, Cursor, JetBrains IDEs
Devin CLI	A near-top score for less per task, and cloud agents	Free, without Fusion; Pro $20; Max $200	61.5, with Opus 5.5 and GPT-6 Luna	Mac, Windows, Linux
GitHub Copilot	An agent and suggestions in almost any IDE	Free; Pro $10; Pro+ $39; Max $100	59.9, with Copilot Fusion at Max, from Pro+	VS Code, Visual Studio, JetBrains IDEs, Eclipse, Xcode; a CLI
Muse Code	The cheapest paid plan	Everyday $5; High $15; Power $50; Maximum $200	58.9, with Muse Wonder Full, from Power	Mac, Windows, Linux
Codex	The lowest cost per task among the top scorers, and a free start	Free and Go ($8) in the desktop app, with GPT-6 Luna; Plus $20; Pro $100, $200 or $500	54.7, with GPT-6 Astra; 54.1, with GPT-6.1 Sol	Mac, Windows, Linux
Google Antigravity	Gemini, with a free start	Free; Google AI Pro $19.99; Ultra $99.99 or $199.99	53.2, with Gemini 4 Argon, from Google AI Pro	Mac, Windows, Linux
OpenCode	Choosing the model, free and open source	Free; models at their providers’ prices; Go $10; Go Plus $40	48.9, with GLM-5.3, the one model tested	Mac, Windows, Linux
Grok Build	People who already pay for Grok or X	Free, limited; Plus $8; Premium $30; Super $100; Ultra $200	46.6, with Grok 4.7	Mac, Windows, Linux
Cursor	Writing code yourself, with suggestions as you type	Free; Plus $8, Grok and Composer only; Premium $30; Super $100; Ultra $200	Not on the index	Mac, Windows, Linux
Enjoy, with six of these	Using more than one in one conversation	A monthly subscription on top of your plans, free for 7 days (pricing)	The score of the agent you pick	Mac, Windows, Linux

Prices are per month as each maker lists them in the US, before tax. The index also scores Mistral Vibe, at 48.3 for $1.26 a task with Mistral Large 4, DeepSeek’s own agent, at 46.3 for $0.30, and Kimi Code’s CLI, at 41.4 with Kimi K3. Cline and Kiro aren’t on it; our Claude Code alternatives and Cursor alternatives guides cover them.

Which to pick

• The hardest work: Claude Code with Opus 5.5.
• The most done for the money: Codex with GPT-6.1 Sol, on ChatGPT Plus.
• You already live in VS Code or a JetBrains IDE: GitHub Copilot, or Claude Code’s or Codex’s own extension.
• You still write much of the code: Cursor, or Copilot.
• The lowest monthly price: Muse Code at $5, once you’ve decided about its Contributor mode.
• No money at all: OpenCode with free or local models, Antigravity’s free quota, or Codex on ChatGPT’s Free plan.
• Code you can read and change: OpenCode, Grok Build or the Codex CLI.
• Windows, with a sandbox: Codex, which “runs natively on Windows using PowerShell and the Windows sandbox.” Claude Code’s sandboxing needs WSL 2.
• No terminal: Claude’s desktop app, the ChatGPT app, the Antigravity app or Enjoy.
• More than one lab, because one plan keeps running out: Enjoy, or two agents side by side.

How the agents score

Agent and model	Index	API cost per task	Time per task	Plans from
Claude Code, Opus 5.5 at max effort	61.7	$11.42	61 minutes	Claude Pro, $20
Devin CLI, Fusion of Opus 5.5 and GPT-6 Luna	61.5	$8.00	51 minutes	Devin Pro, $20
GitHub Copilot, Copilot Fusion at Max	59.9	$5.59	36 minutes	Copilot Pro+, $39
Claude Code, Sonnet 5.5 at max effort	59.3	$13.30	80 minutes	Claude Pro, $20
Muse Code, Muse Wonder Full at max effort	58.9	Not given	35 minutes	Power, $50
Codex, GPT-6 Astra at max effort	54.7	$7.04	25 minutes	ChatGPT Plus, $20
Codex, GPT-6.1 Sol at max effort	54.1	$1.50	24 minutes	ChatGPT Plus, $20
Antigravity CLI, Gemini 4 Argon at high effort	53.2	$11.73	35 minutes	Google AI Pro, $19.99
OpenCode, GLM-5.3 at max effort	48.9	$5.23	56 minutes	Free, plus the model
Codex, GPT-6.1 Terra at extra high effort	48.2	$0.69	16 minutes	ChatGPT Plus, $20
Grok Build, Grok 4.7 at extra high effort	46.6	$9.48	36 minutes	Grok’s plans, from $8

The index scores each pairing of agent and model on 290 tasks from three benchmarks, weighted equally: FrontierCode v1.1, from Cognition, which also makes Devin; Terminal-Bench 4.0, from the Laude Institute; and SWE-Atlas-QnA-AA, from Scale AI, with Artificial Analysis’s own grading. The figures here are as it stood on November 1.

Read it with three things in mind. The costs are “pay-per-token API cost, not consumer plan pricing”: on a plan, a task draws on your limits instead. Its top Gemini result, 59.4 with Gemini 4 Carbon, is left out above because the index marks it unavailable and Google’s models page doesn’t list Carbon for any plan. And a benchmark isn’t your codebase: a few points’ gap is worth less than a week of your own work on each.

Claude Code

Best for: the hardest work, when results matter more than cost.

Anthropic calls Claude Code “an agentic coding tool that reads your codebase, edits files, runs commands, and integrates with your development tools,” available “in your terminal, IDE, desktop app, and browser.” It has the top score on the index, and when a limit stops it partway through a task, its docs say “Claude Code waits in the open session and continues the task on its own after the limit resets.”

Price: Claude’s pricing has Pro at $20 a month, or $17 billed yearly; Max at $100, with “5x Pro usage”; and Ultra at $200, with “10x Pro usage” every week and “No 5-hour limit” for verified accounts. Its setup guide says “The free claude.ai plan does not include Claude Code access.” Our plans guide helps choose.

Watch for: the limits are Claude’s. “These limits are shared across Claude and Claude Code, meaning all activity in both tools counts against the same limits,” says Anthropic’s help center, and of Pro, “The number of messages you can send will vary.” Its top score is slow and dear: an hour and $11.42 a task, against 24 minutes and $1.50 for Codex with Sol. It isn’t open source: its license reads “© Anthropic PBC. All rights reserved.” And its setup guide lists sandboxing on native Windows as “Not supported.”

Devin CLI

Best for: a score close to Claude Code’s for less per task, and handing work to an agent in the cloud.

Cognition’s Devin “runs in the cloud or on your machine, tests in its own browser, and won’t stop until the PR is ready to merge.” The score comes from Fusion in its CLI, which “pairs a frontier lead model with a cost-efficient sidekick”: with Opus 5.5 leading and GPT-6 Luna beside it, Fusion scores 61.5, within a point of Claude Code, for $8.00 a task against $11.42. The CLI installs on macOS, Linux, WSL and Windows, and Devin Desktop, its editor, “is the new name for Windsurf.”

Price: its pricing has Free, with a “Light quota to code with agents”; Pro at $20 a month, with “access to frontier models from OpenAI, Claude, Gemini, and SpaceXAI” and “Access to cloud agents (Devin Cloud)”; and Max at $200, with “Significantly higher quotas.” Teams is $80 a month plus $40 for each full developer seat.

Watch for: Fusion is “not included in free or trial tiers,” and it “bills each model in the pairing at its own rate.” The pricing page puts no number on its quotas: “Each paid plan comes with a usage allowance that refreshes automatically on a daily and weekly basis,” and “Cost per message varies based on the model used.” One of the index’s three benchmarks, FrontierCode, is Cognition’s own.

GitHub Copilot

Best for: an agent and suggestions as you type, in almost any IDE.

Copilot works in VS Code, Visual Studio, JetBrains IDEs, Eclipse and Xcode, per its feature matrix, and “All plans include Copilot CLI and the Copilot app.” Its score comes from Copilot Fusion, “an intelligent runtime model router and orchestrator” that “chooses models and execution patterns for your task, then returns a single response.” At its Max level, Fusion scores 59.9 for $5.59 a task, the lowest cost given for any setup above 55.

Price: Copilot Free costs nothing. Pro is $10 a month, with $15 of GitHub AI Credits; Pro+ is $39, with $70; and Max is $100, with $200.

Watch for: the score isn’t on the $10 plan: GitHub’s table marks Fusion High and Fusion Max “Not included” on Pro, and its plans page leaves Opus 5.5, GPT-6.1 Sol and Gemini 4 Argon out of it too. “Included AI credits do not carry over between months,” says its billing guide, and past them, “additional usage may be capped.” On Free, “access to models is available through auto model selection only.” Our Copilot alternatives guide has more.

Muse Code

Best for: the cheapest paid start, and on Power, a score close to the top.

Muse Code is “Meta’s coding agent for the terminal and CI,” with “approvals and an OS sandbox on from the first run,” per its docs, and it “runs on macOS, Linux, and Windows from one codebase.” With Muse Wonder Full, it scores 58.9 in 35 minutes a task; Artificial Analysis gives no cost for that model. With Muse Spark 1.3, it scores 45.7 for $3.91.

Price: Meta’s Muse Code page lists Everyday Usage at $5 a month, to “Send 10-50 prompts every 5 hours”; High Usage at $15, with “5x more usage”; Power Usage at $50, with “Access to the most powerful Muse models, including Muse Wonder Full” and 20 times Everyday; and Maximum Usage at $200, with 80 times. It lists no free plan.

Watch for: Contributor mode. “Muse Code subscriptions default to Contributor mode. In Contributor mode, you get more usage from your plan in exchange for letting Meta use your content to train future Meta models,” says its subscription guide. You can turn it off, but “Muse Code saves the setting on the computer you’re using,” and “On a new computer, your first interactive session turns Contributor mode on.” The near-top score starts at $50, not $5. Our Muse Code guide has more.

Codex

Best for: the most done for the money, and a free way to start.

OpenAI’s Codex runs in the ChatGPT desktop app, for “macOS, Windows, or Linux,” and in a terminal, your editor and OpenAI’s cloud. The CLI is open source under Apache 2.0. On the index, GPT-6.1 Sol scores 54.1 for $1.50 a task in 24 minutes, and GPT-6.1 Terra at extra high effort scores 48.2 for $0.69.

Price: OpenAI’s pricing gives Free, and Go at $8 a month, “GPT-6 Luna at Standard speed in the desktop app, subject to rollout.” Plus, at $20, is “Power a few focused coding sessions each week,” with “Codex on the web, in the CLI, in the IDE extension, and on iOS.” Pro is $100, $200 or $500, and “Pro plans currently have no five-hour limit.”

Watch for: OpenAI estimates 15 to 160 local messages every five hours with Sol on Plus, but “These estimates are not fixed message limits,” and “Weekly limits may also apply.” The free plans’ model is a long way down: Codex with Luna scores 31.1. “ChatGPT Work and Codex share usage,” and from December OpenAI plans to share that allowance with the rest of ChatGPT. Our Plus vs Pro guide helps choose.

Google Antigravity

Best for: Gemini, with a free start, in Google’s own app or terminal.

The Antigravity App is Google’s “dedicated platform to work with agents,” with a CLI, agy, beside it, and downloads for macOS, Windows, Linux and Googlebook. With Gemini 4 Argon at high effort, the CLI scores 53.2 for $11.73 a task. The nearest thing on the index to the free plan’s model is Antigravity’s SDK with Gemini 3.8 Flash, at 33.3.

Price: free, with “Basic weekly rate limits,” per its pricing. Google AI Pro is $19.99 a month, and Ultra is $99.99 or $199.99. Gemini 4 Argon starts with Pro, and Google’s models page says the Claude models there are “Available on Google AI Pro for non-trial subscriptions only.”

Watch for: the free plan’s Claude is going. Claude Sonnet 4.6 and Opus 4.6, the only Claude models on free accounts, are marked “Will be removed on November 2, 2026,” as our memo explains. The free quota is “A meaningful quota, refreshed weekly,” and “Usage limits for this service are subject to modification.” For personal accounts, Google’s terms say it uses your interactions “to evaluate, develop, and improve Google and Alphabet research, products, services and machine learning technologies,” and that “Google employees and contractors may access, view, review and use Interactions,” unless you change that in settings. “There is currently no support” for bringing your own key. Our Antigravity alternatives guide has more.

OpenCode

Best for: choosing the model, free and open source.

OpenCode is “The open source AI coding agent,” under the MIT license on GitHub, with more than 223,000 stars. Its homepage lists “75+ LLM providers through Models.dev, including local models,” and you can “Log in with OpenAI to use your ChatGPT Plus or Pro account” or “Log in with GitHub to use your Copilot account.” The index has tested it with one model, GLM-5.3, at 48.9 for $5.23 a task, so that score says as much about the model as the agent. Our local models guide shows how to run it with no bill at all.

Price: free, with “a set of free models that you can use without creating an account.” Other providers bill at their own prices, and OpenCode Go is $10 a month, or $40 for Go Plus.

Watch for: your Claude plan doesn’t work in it: “There are plugins that allow you to use your Claude Pro/Max models with OpenCode. Anthropic explicitly prohibits this,” say its provider docs. Out of the box it asks little: “Most permissions default to "allow",” per its permissions page. Its free models are free “for a limited time,” and of one, “The team is using this time to collect feedback and improve the model.”

Grok Build

Best for: people who already pay for Grok or X, and an open-source agent from $8.

Grok Build is “a powerful and extensible coding agent,” its docs say, and its code is open under Apache 2.0, with more than 27,000 stars. It installs with one command on Mac and Linux, or a PowerShell script on Windows. With Grok 4.7 at extra high effort, it scores 46.6 for $9.48 a task.

Price: Grok’s pricing lists Free, where Grok Build is “Limited”; Plus at $8 a month, with “Grok Chat, Voice, Imagine, and Build” and Cursor; Premium at $30, with four times Plus’s usage; Super at $100; and Ultra at $200.

Watch for: it’s the lowest score here, at one of the higher costs per task. Usage is one pool across Grok’s products, so chats and images draw on what Grok Build has left, and Premium lists “Latest Grok models” among what it adds to Plus, so check which models Plus gets. And its sandbox is “Off by default,” with “Landlock on Linux, Seatbelt on macOS,” and none named for Windows. Our Grok Build guide has more.

Cursor

Best for: writing code yourself, with an agent beside you for the bigger jobs.

Cursor is an AI code editor whose desktop app “is available for macOS, Windows, and Linux.” Tab “is Cursor’s AI-powered autocomplete. It suggests code as you type,” which Claude Code doesn’t do. Artificial Analysis’s index lists no Cursor setup, so we give it no score.

Price: its pricing has Free, with “Limited agent requests”; Plus at $8 a month, where “Only Grok models are available on this plan”; Premium at $30, with “4x” Plus’s usage; Super at $100; and Ultra at $200.

Watch for: Claude, GPT and Gemini cost extra. “Third-party models like Claude, GPT, and Gemini aren’t part of the included pool,” says its models page, and “On Premium and above, they’re available in Cursor at the model’s API list price from the first request.” That page lists GPT-5.6, not OpenAI’s GPT-6 models, and OpenAI has proposed November 12 as the end of its models there. You can keep the editor and let Claude Code’s or Codex’s extension do the agent’s work on your own plan, as our Cursor vs Claude Code guide shows.

Enjoy, with six of these

Best for: using more than one agent, in one conversation, when one plan keeps running out.

Enjoy isn’t a tenth agent, and it adds no score of its own. It’s a desktop app for Mac, Windows and Linux that runs Claude Code, Codex, Muse Code, Grok Build, Antigravity, and local models through OpenCode, on the plans you already have. It doesn’t run Copilot, Devin or Cursor. What it adds:

• Switch agents mid-conversation. Hand a job on at any point: “The new agent gets the conversation so far, including replies from other agents, before it works on your message.” Plan with Claude Code, then build with Codex.
• Keep going through limits. When an agent runs out, Enjoy switches to another of your accounts with usage left, or waits for the reset and carries on. Select Stop waiting to hand the conversation to another agent instead.
• Every project at once. Each conversation runs its own agent, and JOY keeps an eye on all your projects: ask it what needs you, or start work in any of them.
• Usage in view. It shows what each signed-in account has left and when each limit resets.

Price: Enjoy is one monthly subscription, and the first seven days are free, with a card to start. It includes your phone and any browser and your other computers, and each teammate you invite needs a subscription of their own. See pricing.

Watch for: with one agent, you may not need it. Enjoy doesn’t change any lab’s limits or add one lab’s models to another’s plan, and it runs the Codex CLI, which OpenAI includes from Plus. On Antigravity, be careful. Enjoy runs Google’s agy CLI in its headless mode, and Google’s FAQ says that mode “is intended for direct scripting and CI workflows,” and that “Wrapping agy inside third-party coding agents or applications, or automating account rotation, is not supported under consumer plans.” Its terms add that using “third party software, tools, or services to access the Service” “may be grounds for suspension or termination,” and Google hasn’t said how it applies that to an app like Enjoy. If that’s a risk you won’t take with your Google account, use Antigravity in Google’s own app, and the other agents in Enjoy. In its default mode, Antigravity in Enjoy also approves every tool. Enjoy runs Claude Code headless too, which Anthropic’s terms allow for “an end user” signing in “to the unmodified Claude Code binary with their own Claude subscription,” but if the reports in issue #97074, still open, are right, it uses the five-hour window faster than in a terminal. Waiting for a reset needs “Your computer must be on, with Enjoy open, when the usage resets,” and scheduled workflows run only while it’s awake with Enjoy open. Enjoy isn’t open source, and it’s a monthly subscription on top of your AI plans once the free week ends.

When to stay with the agent you have

If your agent does your work and you rarely run out, none of this is a reason to move. The gaps near the top are small: Claude Code, Devin, Copilot and Muse Code all sit within three points of each other, and how well you know an agent’s habits counts for more than that. If a limit is the problem, waiting, credits or a bigger plan may cost less than moving, as our guides for Claude Code and Codex work out.

Move when the work is hard enough that another agent’s results show up in your reviews, when the price is what stops you, or when you need something yours lacks, such as a free plan, open code or Windows sandboxing. And you don’t have to pick only one: Claude Pro and ChatGPT Plus together cost $40 a month, with limits that run out separately. One AGENTS.md can instruct both, as Claude Code’s and Codex’s docs describe, and our guide to using both covers the free ways to do it.

How we compared

We read each maker’s homepage, pricing, plans and docs, the GitHub repositories of the open-source agents, and Artificial Analysis’s index and its Claude Code vs Antigravity CLI comparison on November 1, 2026, and every one is linked below. Licenses and star counts come from GitHub that day, and every score, cost and time is as the index showed it that day. OpenAI’s help center wouldn’t load for us, so what we say about December’s shared allowance and the November 12 proposal comes from our own memos, linked where we say it. We quote each maker’s own words for its limits and leave out what it hasn’t documented. We didn’t run these agents against each other ourselves: the scores are Artificial Analysis’s, and they’re a benchmark, not a verdict on your code. This category changes weekly, so we recheck this guide against its sources at least once a month, and the date at the foot of the page says when we last did.