AI coding tools now do much more than complete a line of code. They can inspect a repository, edit multiple files, run tests, open pull requests, and delegate work to parallel agents. Their feature lists increasingly overlap, but the practical differences are still substantial: where the work runs, how permissions are controlled, and how well the product fits the development system you already use.
This guide compares OpenAI Codex, Anthropic Claude Code, Google Antigravity, and GitHub Copilot CLI. It is not a model leaderboard. Vendor benchmarks use different models, harnesses, settings, and test conditions, so combining them into a universal ranking would be misleading.
The comparison at a glance
| Tool | Main interfaces | Distinctive workflow | Strong reason to choose it |
|---|---|---|---|
| Codex | Desktop app, CLI, IDE, cloud | Parallel agents, worktrees, Skills, cloud tasks | You already use ChatGPT and want local and cloud work to connect |
| Claude Code | Terminal, IDE, web or desktop cloud sessions | Long-running tasks, subagents, hooks, Artifacts, new Projects | You need large-task decomposition and visible team coordination |
| Antigravity | CLI, desktop IDE, editor integrations | Shared Antigravity 2.0 harness and asynchronous work | You used Gemini CLI or work primarily in Google's ecosystem |
| GitHub Copilot CLI | Terminal, GitHub, IDE | Plan mode, Autopilot, GitHub actions, local or cloud sandbox | Your issues, pull requests, reviews, and access controls live in GitHub |
These are product positions, not performance guarantees. Results can change substantially with the selected model, permissions, instructions, and repository environment.
Codex: the broadest bridge between local and cloud work
Codex is available through a desktop app, CLI, IDE extension, and Codex cloud. Cloud tasks run in isolated environments and can continue in the background before you review the summary and diff. The desktop app can supervise parallel agents and uses Git worktrees to isolate simultaneous tasks in the same repository.
Skills, Plugins, Sites, and annotations have also expanded Codex beyond software development into research, data analysis, presentations, and internal tools. If your current subscription includes Codex and you move between a Windows or Mac desktop, the terminal, an IDE, and GitHub, Codex offers a relatively coherent starting point.
The caveat is that features are spread across several surfaces. Enterprise previews may not appear in an individual account, and the Agents API is a developer platform for building an agent into another product, not a requirement for using Codex itself.
Related: Codex features in 2026
Claude Code: strong coordination for large bodies of work
Claude Code began as a terminal-first product and now spans IDE, web, desktop, and cloud sessions. It supports subagents, hooks, background tasks, checkpoints, and Artifacts that turn session context into an interactive page for authenticated members of an organization.
On September 17, 2026, Anthropic announced a redesigned Projects beta. After the user defines a goal and repository context, a coordinator can split the work into threads, delegate tasks, review outputs, and assemble the result. That is closer to coordinated project execution than simply opening several independent chat windows.
Availability matters. The first rollout targets selected Pro and Max subscribers who meet the cloud-session and existing-project conditions. Team and Enterprise availability follows later, so an announcement does not mean every account has the feature immediately.
Related: How the redesigned Claude Code Projects work
Antigravity: the new route for individual Gemini CLI users
Google has moved the individual terminal-agent experience from Gemini CLI to Antigravity CLI. Its official announcement states that Gemini CLI stopped serving Google AI Pro, Ultra, and free individual accounts on June 18, 2026. Enterprise licenses and supported API-key access to Gemini CLI continue.
Antigravity CLI and Antigravity 2.0 share the same agent harness and settings. Google positions the CLI as the lightweight terminal interface, with asynchronous background workflows, Skills, Hooks, Subagents, and Plugins. This is more than a rename because account eligibility, installation, and some configuration paths changed.
Related: What the Gemini CLI to Antigravity transition means
GitHub Copilot CLI: the shortest path for GitHub-centered teams
GitHub Copilot CLI can edit a project, execute commands, and interact with GitHub from the terminal. It provides plan mode, an autonomous Autopilot mode, and local or cloud sandboxes that create clearer boundaries around file, network, and system access.
For a team whose issues, pull requests, reviews, and permissions already live in GitHub, that integration can remove an extra layer of setup. If the workflow is not centered on GitHub, the advantage is less decisive.
Autopilot is not a reason to approve everything. GitHub explicitly warns that broad tool approval lets the agent run the same commands as the current user. Starting with a plan, restricted paths, and a sandbox is more defensible than enabling unrestricted execution on the first task.
Which one should you choose?
You already pay for ChatGPT and need coding plus general work
Start with Codex. Its local, cloud, and non-coding workflows are connected closely enough that you may not need another subscription for the first trial.
You maintain a large or multi-repository system
Test Claude Code and Codex on the same real issue. Claude's redesigned Projects emphasize coordinator-led threads; the Codex app emphasizes supervising parallel work in isolated worktrees. Measure passing tests, unrelated edits, review time, and the amount of manual cleanup rather than relying on a leaderboard.
You previously relied on Gemini CLI
Individual users should evaluate Antigravity CLI instead of repeatedly troubleshooting old Gemini CLI OAuth instructions. Enterprise and API-key users should verify their existing entitlement before deciding whether to migrate.
Your team operates almost entirely in GitHub
Copilot CLI offers the lowest integration overhead when issue tracking, pull requests, review, and sandbox policies should remain in GitHub.
Run the same evaluation task
Do not evaluate an agent by asking it to explain a repository. Give each candidate the same moderate, real task: reproduce and fix a bug, add tests, or update an outdated dependency. Record:
- Whether the first change passes the test suite.
- How many unrelated files it modifies.
- Whether permission prompts are understandable.
- Whether interrupted work resumes cleanly.
- How long human review and correction take.
- The plan allowance or API cost consumed by the task.
The best product is not the one with the most dramatic demo. It is the one that completes work reliably in your repository without moving a large cleanup burden into review.
Permissions matter more than model rankings
An agent that can run a shell, read environment variables, access the network, or open a pull request has a larger blast radius than a chatbot. Keep production credentials out of test environments, avoid permanent blanket approvals, and use separate branches or worktrees, least privilege, sandboxes, and recoverable version control with every product in this comparison.
Official sources
- OpenAI: Introducing the Codex app
- OpenAI: Codex cloud
- Anthropic: Claude Code product and plans
- Anthropic: Redesigned Claude Code Projects
- Google: Introducing Antigravity CLI
- Google: Gemini CLI transition announcement
- GitHub: About Copilot CLI
- GitHub: Copilot CLI Autopilot
繁體中文： AI 程式開發工具比較 2026