Compare · checked 2 Nov 2026

AI coding agents compared

The AI agents that write and change code for you: where each runs, what it does without asking, what it asks first, who can use it and what it costs.

Facts come from each company's own pages, with the date we checked. We don't rank them, and we don't claim to have tested them: whether coding agents make developers faster is a question for the evidence, which our rated use cases will grade.

See them side by side Plan a rollout

Side by side

	Alibaba (Qwen team)Qwen Code	Amazon Web ServicesKiro	AnthropicClaude Code	CursorCursor	GitHub (Microsoft)GitHub Copilot cloud agent and agent mode	GoogleGoogle Antigravity	OpenAIOpenAI Codex
Status	Not stated	Generally available since 17 November 2025; mobile app listed as preview. source ↗	Generally available since 22 May 2025. source ↗	Not stated	Coding agent generally available since 25 September 2025; renamed Copilot cloud agent on 1 April 2026. source ↗	Not stated	Not stated
Who can use it	Anyone connecting a provider: QwenCloud, Alibaba ModelStudio, third-party providers or a local server. The Qwen OAuth free tier ended 15 April 2026. source ↗	Individuals on Free or paid plans; teams with central admin, including AWS IAM Identity Center. source ↗	Pro, Max, Ultra, Team and Enterprise plans, or Anthropic Console (API) accounts; not the Free plan. source ↗	Free, Plus, Premium, Super, Ultra, Teams and Enterprise plans; cloud agents need Premium or higher, Teams or Enterprise. source ↗	Cloud agent: paid plans (Pro, Pro+, Max, Business, Enterprise; admins enable for Business and Enterprise). IDE agent mode: every plan, including Free. source ↗	Individuals free with a Google account; higher limits with Google AI Pro or Ultra; organisations through Google Cloud. source ↗	ChatGPT Free, Go, Plus, Pro, Business, Edu and Enterprise; or an API key (CLI, SDK and IDE only, no cloud features). source ↗
Price	Client is free; model use is paid to the provider: usage-based Token Plan, monthly-credit Token Plan or fixed-fee Coding Plan. No free tier. source ↗	Free (50 credits), Pro $20 (1,000), Pro+ $40, Pro Max $100, Power $200 a month; extra credits $0.04 each; Enterprise by quote. source ↗	Pro $20/month, Max $100, Ultra $200; Team from $20/seat; Enterprise seat plus API-rate usage. Limits reset every 5 hours and weekly. source ↗	Free; Plus $8, Premium $30, Super $100, Ultra $200 a month; Teams from $30/user. Usage refreshes weekly; top-ups and third-party models extra. source ↗	Free $0; Pro $10, Pro+ $39, Max $100 a month; Business and Enterprise via sales. Use draws GitHub AI Credits ($0.01 each). source ↗	Free with weekly limits; Google AI Pro and Ultra add higher limits and AI credits; Business from $30/seat/month or pay-as-you-go. source ↗	Go $8, Plus $20, Pro from $100 a month; Business $20/user (annual); Enterprise by quote. API key: usage-based. Limits vary by task. source ↗
Where it runs	Terminal on your machine (macOS, Linux, Windows); VS Code extension in beta. source ↗	Desktop IDE and terminal CLI on your machine; Kiro Web runs sandboxed cloud sessions in the browser; mobile app. source ↗	Terminal, VS Code and JetBrains, desktop app, browser and mobile; cloud sessions run in Anthropic-managed virtual machines. source ↗	Desktop editor and CLI on your machine; web (cursor.com/agents), iOS, Slack, GitHub and Linear; cloud agents in isolated cloud VMs. source ↗	Cloud agent: GitHub Actions, a cloud sandbox or your own remote environment. Agent mode: VS Code, Visual Studio, JetBrains, Xcode, Eclipse. source ↗	Desktop app and IDE (macOS, Windows, Linux) and terminal CLI (agy) on your machine; browser automation in a sandboxed Chromium. source ↗	Terminal (CLI), IDE extension, ChatGPT desktop, web and mobile apps; cloud tasks run in isolated hosted workspaces. source ↗
Models	Qwen models plus OpenAI, Anthropic, Gemini, GLM, Kimi, MiniMax and local models (Ollama, vLLM); you choose and can switch providers mid-session. source ↗	Anthropic Claude, OpenAI GPT and open-weight models (DeepSeek, Qwen, MiniMax, GLM, Kimi); pick one or Auto, each with a credit multiplier. source ↗	Anthropic Claude models only (Opus, Sonnet, Haiku, Fable); default Opus 5.5; switch with /model. Admins can restrict choices. source ↗	You pick a model per task from supported frontier models; the Plus plan is limited to Grok models. source ↗	Models from OpenAI, Anthropic, Google, Microsoft and others; selectable or Auto, limited by plan and organisation policy. source ↗	You choose: Gemini 4 Argon, Gemini 3.8 Flash, Gemini 3.1 Pro, Claude Sonnet 5.5, Claude Opus 5.5, GPT-OSS 120B; availability depends on plan. source ↗	OpenAI models such as Astra, GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Luna; choose with /model or --model; options vary by plan. source ↗
Does without asking	Read-only operations run without asking. Edits run unasked only in Auto-Edit, Auto or YOLO modes; YOLO also runs every shell command. source ↗	By default reads any workspace file and runs read-only git and system-info commands without asking. source ↗	Reads files and runs built-in read-only commands (ls, grep, read-only git) without asking in every mode. source ↗	Edits code, runs terminal commands, searches files and the web, controls a browser; cloud agents never ask for approval and push pull requests. source ↗	Cloud agent researches, plans, edits files, runs tests and linters, pushes to one branch and can open a draft pull request. source ↗	Default preset edits workspace and temp files and runs commands inside its sandbox without prompting. source ↗	In the Auto preset it reads files, edits and runs commands in the working directory; network is off by default. source ↗
Asks first	Five modes: Plan, Ask Permissions, Auto-Edit, Auto (LLM classifier) and YOLO. Docs conflict on whether Ask Permissions or Auto is the default. source ↗	File edits and other shell commands prompt by default; .git writes always prompt. Rules can allow, ask or deny; IDE offers Autopilot or Supervised. source ↗	Default (Manual) asks before edits, most shell commands and web access. Other modes: acceptEdits, plan, auto (classifier), dontAsk, bypassPermissions. source ↗	Default Auto-review: allowlisted calls run, other commands go to a sandbox or a classifier. Other modes: Allowlist, Run Everything. source ↗	Workflows wait for human approval; humans must review and merge its PRs; only write-access users can trigger it. IDE agent mode asks before terminal commands. source ↗	Default asks before commands outside the sandbox, edits outside the workspace, web and MCP actions. Presets: Request Review, Default, Auto (preview), Turbo. source ↗	Asks to edit outside the workspace or use the network. Modes: Ask for approval (recommended), Approve for me, Full access, Custom. source ↗
Your code and training	Qwen Code itself does not use prompts, code or responses for training; your provider's terms apply. Usage statistics on by default, can be disabled. source ↗	Free and individual users' content may be used for service improvement, including model training, unless they opt out. Enterprise content is not used. source ↗	Free/Pro/Max/Ultra: trained on only if you allow (5-year retention; else 30 days). Team, Enterprise, API: not trained on; 30-day retention. source ↗	Privacy Mode on: no training, zero retention with providers. Off: Cursor may use code and prompts to train models. On by default for teams. source ↗	Free, Pro, Pro+ and Max interaction data may be used for training unless you opt out; Business and Enterprise data is not used. source ↗	Enterprise prompts, code and transcripts are never used to train models. Individual-account training terms not stated on the pages checked. source ↗	Business: no training on business data by default. API-key use follows the API organisation's data settings. Personal-plan policy not opened. source ↗
Open source	Yes, Apache-2.0. source ↗	No. The public GitHub repository is an issue and feedback tracker; product source code is not hosted there. source ↗	No. Licence notice: all rights reserved, use subject to Anthropic's Commercial Terms. The repository holds plugins and examples. source ↗	No source or licence published; the public GitHub repository holds only issue templates, a README and a security policy. source ↗	The VS Code Copilot Chat extension is MIT-licensed, moved into the VS Code repository in May 2026. No licence stated for the cloud agent. source ↗	Not stated for Antigravity. The earlier Gemini CLI is Apache-2.0 open source, now aimed at enterprise Gemini Code Assist. source ↗	Codex CLI repository is Apache-2.0. source ↗
Launched	Announced and open-sourced with Qwen3-Coder on 22 July 2025, described then as a research-purpose CLI tool. source ↗	Introduced in preview on 14 July 2025, free during preview with limits. source ↗	24 February 2025, as a limited research preview alongside Claude 3.7 Sonnet. source ↗	Not stated	Not stated	Not stated	Not stated

Scroll sideways on small screens. Every fact links to the company page that states it.

Each one in detail

Alibaba (Qwen team)

Qwen Code

Open-source terminal coding agent from Alibaba's Qwen team, originally based on Gemini CLI, that works with Qwen and other providers' models through your own plan or API key.

Status: Not stated
Who can use it: Anyone connecting a provider: QwenCloud, Alibaba ModelStudio, third-party providers or a local server. The Qwen OAuth free tier ended 15 April 2026. source ↗
Price: Client is free; model use is paid to the provider: usage-based Token Plan, monthly-credit Token Plan or fixed-fee Coding Plan. No free tier. source ↗
Where it runs: Terminal on your machine (macOS, Linux, Windows); VS Code extension in beta. source ↗
Models: Qwen models plus OpenAI, Anthropic, Gemini, GLM, Kimi, MiniMax and local models (Ollama, vLLM); you choose and can switch providers mid-session. source ↗
Does without asking: Read-only operations run without asking. Edits run unasked only in Auto-Edit, Auto or YOLO modes; YOLO also runs every shell command. source ↗
Asks first: Five modes: Plan, Ask Permissions, Auto-Edit, Auto (LLM classifier) and YOLO. Docs conflict on whether Ask Permissions or Auto is the default. source ↗
Your code and training: Qwen Code itself does not use prompts, code or responses for training; your provider's terms apply. Usage statistics on by default, can be disabled. source ↗
Open source: Yes, Apache-2.0. source ↗
Launched: Announced and open-sourced with Qwen3-Coder on 22 July 2025, described then as a research-purpose CLI tool. source ↗

Release notes ↗

Amazon Web Services

Kiro

AWS's agentic coding product with a desktop IDE, terminal CLI, browser-based cloud sessions and mobile app, billed in credits; AWS points Amazon Q Developer IDE users to it.

Status: Generally available since 17 November 2025; mobile app listed as preview. source ↗
Who can use it: Individuals on Free or paid plans; teams with central admin, including AWS IAM Identity Center. source ↗
Price: Free (50 credits), Pro $20 (1,000), Pro+ $40, Pro Max $100, Power $200 a month; extra credits $0.04 each; Enterprise by quote. source ↗
Where it runs: Desktop IDE and terminal CLI on your machine; Kiro Web runs sandboxed cloud sessions in the browser; mobile app. source ↗
Models: Anthropic Claude, OpenAI GPT and open-weight models (DeepSeek, Qwen, MiniMax, GLM, Kimi); pick one or Auto, each with a credit multiplier. source ↗
Does without asking: By default reads any workspace file and runs read-only git and system-info commands without asking. source ↗
Asks first: File edits and other shell commands prompt by default; .git writes always prompt. Rules can allow, ask or deny; IDE offers Autopilot or Supervised. source ↗
Your code and training: Free and individual users' content may be used for service improvement, including model training, unless they opt out. Enterprise content is not used. source ↗
Open source: No. The public GitHub repository is an issue and feedback tracker; product source code is not hosted there. source ↗
Launched: Introduced in preview on 14 July 2025, free during preview with limits. source ↗

Release notes ↗

Anthropic

Claude Code

Anthropic's coding agent that reads a codebase, edits files, runs commands and opens pull requests, from the terminal, IDEs, a desktop app, the browser or cloud sessions.

Status: Generally available since 22 May 2025. source ↗
Who can use it: Pro, Max, Ultra, Team and Enterprise plans, or Anthropic Console (API) accounts; not the Free plan. source ↗
Price: Pro $20/month, Max $100, Ultra $200; Team from $20/seat; Enterprise seat plus API-rate usage. Limits reset every 5 hours and weekly. source ↗
Where it runs: Terminal, VS Code and JetBrains, desktop app, browser and mobile; cloud sessions run in Anthropic-managed virtual machines. source ↗
Models: Anthropic Claude models only (Opus, Sonnet, Haiku, Fable); default Opus 5.5; switch with /model. Admins can restrict choices. source ↗
Does without asking: Reads files and runs built-in read-only commands (ls, grep, read-only git) without asking in every mode. source ↗
Asks first: Default (Manual) asks before edits, most shell commands and web access. Other modes: acceptEdits, plan, auto (classifier), dontAsk, bypassPermissions. source ↗
Your code and training: Free/Pro/Max/Ultra: trained on only if you allow (5-year retention; else 30 days). Team, Enterprise, API: not trained on; 30-day retention. source ↗
Open source: No. Licence notice: all rights reserved, use subject to Anthropic's Commercial Terms. The repository holds plugins and examples. source ↗
Launched: 24 February 2025, as a limited research preview alongside Claude 3.7 Sonnet. source ↗

Release notes ↗

Cursor

AI code editor whose Agent edits code and runs commands locally, plus a CLI, web and iOS access, and cloud agents that work in isolated VMs.

Status: Not stated
Who can use it: Free, Plus, Premium, Super, Ultra, Teams and Enterprise plans; cloud agents need Premium or higher, Teams or Enterprise. source ↗
Price: Free; Plus $8, Premium $30, Super $100, Ultra $200 a month; Teams from $30/user. Usage refreshes weekly; top-ups and third-party models extra. source ↗
Where it runs: Desktop editor and CLI on your machine; web (cursor.com/agents), iOS, Slack, GitHub and Linear; cloud agents in isolated cloud VMs. source ↗
Models: You pick a model per task from supported frontier models; the Plus plan is limited to Grok models. source ↗
Does without asking: Edits code, runs terminal commands, searches files and the web, controls a browser; cloud agents never ask for approval and push pull requests. source ↗
Asks first: Default Auto-review: allowlisted calls run, other commands go to a sandbox or a classifier. Other modes: Allowlist, Run Everything. source ↗
Your code and training: Privacy Mode on: no training, zero retention with providers. Off: Cursor may use code and prompts to train models. On by default for teams. source ↗
Open source: No source or licence published; the public GitHub repository holds only issue templates, a README and a security policy. source ↗
Launched: Not stated

Release notes ↗

GitHub (Microsoft)

GitHub Copilot cloud agent and agent mode

GitHub Copilot's cloud agent (formerly coding agent) works on a repository branch in the cloud and opens draft pull requests; agent mode does multi-step edits inside the IDE.

Status: Coding agent generally available since 25 September 2025; renamed Copilot cloud agent on 1 April 2026. source ↗
Who can use it: Cloud agent: paid plans (Pro, Pro+, Max, Business, Enterprise; admins enable for Business and Enterprise). IDE agent mode: every plan, including Free. source ↗
Price: Free $0; Pro $10, Pro+ $39, Max $100 a month; Business and Enterprise via sales. Use draws GitHub AI Credits ($0.01 each). source ↗
Where it runs: Cloud agent: GitHub Actions, a cloud sandbox or your own remote environment. Agent mode: VS Code, Visual Studio, JetBrains, Xcode, Eclipse. source ↗
Models: Models from OpenAI, Anthropic, Google, Microsoft and others; selectable or Auto, limited by plan and organisation policy. source ↗
Does without asking: Cloud agent researches, plans, edits files, runs tests and linters, pushes to one branch and can open a draft pull request. source ↗
Asks first: Workflows wait for human approval; humans must review and merge its PRs; only write-access users can trigger it. IDE agent mode asks before terminal commands. source ↗
Your code and training: Free, Pro, Pro+ and Max interaction data may be used for training unless you opt out; Business and Enterprise data is not used. source ↗
Open source: The VS Code Copilot Chat extension is MIT-licensed, moved into the VS Code repository in May 2026. No licence stated for the cloud agent. source ↗
Launched: Not stated

Release notes ↗

Google

Google Antigravity

Google's agentic development platform with a desktop app, IDE and terminal CLI; it replaced Gemini CLI for unpaid users in June 2026. Jules is a separate experimental agent.

Status: Not stated
Who can use it: Individuals free with a Google account; higher limits with Google AI Pro or Ultra; organisations through Google Cloud. source ↗
Price: Free with weekly limits; Google AI Pro and Ultra add higher limits and AI credits; Business from $30/seat/month or pay-as-you-go. source ↗
Where it runs: Desktop app and IDE (macOS, Windows, Linux) and terminal CLI (agy) on your machine; browser automation in a sandboxed Chromium. source ↗
Models: You choose: Gemini 4 Argon, Gemini 3.8 Flash, Gemini 3.1 Pro, Claude Sonnet 5.5, Claude Opus 5.5, GPT-OSS 120B; availability depends on plan. source ↗
Does without asking: Default preset edits workspace and temp files and runs commands inside its sandbox without prompting. source ↗
Asks first: Default asks before commands outside the sandbox, edits outside the workspace, web and MCP actions. Presets: Request Review, Default, Auto (preview), Turbo. source ↗
Your code and training: Enterprise prompts, code and transcripts are never used to train models. Individual-account training terms not stated on the pages checked. source ↗
Open source: Not stated for Antigravity. The earlier Gemini CLI is Apache-2.0 open source, now aimed at enterprise Gemini Code Assist. source ↗
Launched: Not stated

Release notes ↗

OpenAI

OpenAI Codex

OpenAI's coding agent, available in the ChatGPT apps, an open-source CLI, an IDE extension and cloud workspaces, included in ChatGPT plans or billed via an API key.

Status: Not stated
Who can use it: ChatGPT Free, Go, Plus, Pro, Business, Edu and Enterprise; or an API key (CLI, SDK and IDE only, no cloud features). source ↗
Price: Go $8, Plus $20, Pro from $100 a month; Business $20/user (annual); Enterprise by quote. API key: usage-based. Limits vary by task. source ↗
Where it runs: Terminal (CLI), IDE extension, ChatGPT desktop, web and mobile apps; cloud tasks run in isolated hosted workspaces. source ↗
Models: OpenAI models such as Astra, GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Luna; choose with /model or --model; options vary by plan. source ↗
Does without asking: In the Auto preset it reads files, edits and runs commands in the working directory; network is off by default. source ↗
Asks first: Asks to edit outside the workspace or use the network. Modes: Ask for approval (recommended), Approve for me, Full access, Custom. source ↗
Your code and training: Business: no training on business data by default. API-key use follows the API organisation's data settings. Personal-plan policy not opened. source ↗
Open source: Codex CLI repository is Apache-2.0. source ↗
Launched: Not stated

Release notes ↗

How we compare

Editorial standards →

The companies' own wordsEach fact comes from the company's product, pricing, help or release-notes pages, checked on 2 Nov 2026. What a company doesn't state, we mark as not stated.
Same rows for everyoneProducts are in alphabetical order by company, with the same questions answered for each. Nobody pays to be listed.
Facts, not a verdictThis page says what each product is and does. Whether it helps your work is a question for evidence and for your own pilot.

Products change fast: check the release notes before you decide. Want to learn first? Free courses. Disclosure: AI Adoption uses AI tools in its research and editorial work, including Claude from Anthropic, which makes one of the products compared here. No company has any say in this page.