You installed the tool, opened a project, and now you have questions. This page answers the ones developers actually ask — from “which model should I use?” to “why did my included usage run out sooner than expected?”
Cursor is an AI-native IDE (fork of VS Code) with inline editing, tab completions, and agent mode built in. Claude Code is a CLI-first agent that also spans IDE, desktop, browser, and Slack. Codex is OpenAI’s multi-surface agent inside the ChatGPT desktop app, CLI, IDE extension, and cloud service with GitHub/Slack/Linear integrations.
Pick Cursor if you want an editor-first Agents Window and Tab workflow. Pick Claude Code for Anthropic’s agent across terminal, IDE, desktop, browser, and Routines. Pick Codex for its shared ChatGPT desktop, CLI, IDE, and Cloud workflow; isolated desktop tasks can use worktrees when selected.
| Aspect | Cursor | Claude Code | Codex |
|---|---|---|---|
| Interface | Agents Window + VS Code fork + web/iOS/CLI | Terminal + IDE + desktop + browser | ChatGPT desktop + CLI + IDE + Cloud |
| Primary model | Configurable (Auto, Grok 4.5, Composer 2.5, Claude, GPT-5.6, etc.) | Account/provider-specific Claude default; Fable 5 available | GPT-5.6 Sol / Terra / Luna by plan and workload |
| Context window | Model-dependent, up to 1M | 1M on current Sonnet / Opus / Fable tiers | Model- and surface-dependent; 1.05M in the GPT-5.6 API |
| Sandboxing | Agent/worktree/cloud sandboxes | Native sandbox + permission modes + web sandboxes | Surface-specific sandboxing; optional worktrees for isolated desktop tasks |
| Cloud execution | Cloud Agents + Automations | Claude Code on the web + managed agents/Routines | Cloud agents (native) |
| MCP support | Yes | Yes | Yes |
| Skills support | Yes | Yes | Yes |
| Best for | Editor-first iteration and Tab | Anthropic-native terminal and multi-surface agents | Codex/ChatGPT multi-surface delivery |
| Starting price | Free tier; paid plans from $20/month | $20/month (Pro) | Included usage by ChatGPT plan, then optional token credits |
For maximum Claude capability: Claude Fable 5 — the tier above Opus at $10/$50 per MTok. It returned globally July 1 and requires usage credits after July 7.
For Claude Code: The default depends on the account/provider. Pro, Team Standard, and Enterprise subscription seats default to Sonnet 5; Max, Team Premium, Enterprise pay-as-you-go, and Anthropic API sessions default to Opus 4.8. Organization policy can override the mapping, and managed-cloud platforms use other defaults. Sonnet 5 has native 1M context and introductory $2/$10 pricing through August 31.
For premium second-pass Claude work: Claude Opus 4.8 — strong architecture, debugging, and review when Sonnet 5 does not meet the quality bar.
For Codex users: choose GPT-5.6 Sol for frontier work, Terra for the everyday balance, or Luna for efficient high-volume work. All three expose a 1.05M API context and 128K max output; a product surface may expose a smaller working context.
For long-running Cursor work: Grok 4.5 — available across desktop, web, iOS, CLI, and SDK at $2/$6 per MTok (fast: $4/$18). It does not replace Composer 2.5; use Composer for smaller speed-focused loops.
For speed-critical work in Cursor: Cursor Composer 2.5 — a smaller coding specialist at $0.50/$2.50 per 1M tokens. Artificial Analysis currently measures Composer + Cursor CLI at 52 on Coding Agent Index v1.1, versus 76 for Grok 4.5 + Grok Build and 80 for GPT-5.6 Sol + Codex; harness differences are part of those results.
For extreme context or multimodal: Gemini 3.1 Pro — 1M token context, strong image/video analysis, and Deep Think mode.
See the full Model Comparison for detailed specs and pricing.
Individual Plans
Team Plans
Claude Fable 5: Global access returned July 1. The temporary included-usage window ended July 7; further use on eligible plans requires usage credits.
See the full Pricing Calculator for detailed cost scenarios.
Absolutely. Many developers use Cursor for interactive editing and tab completions, Claude Code for terminal-based autonomous tasks, and Codex for cloud-based automation and GitHub issue handling. The tools are complementary, not mutually exclusive. Your project configuration files (.cursor/rules/, CLAUDE.md, AGENTS.md) can coexist in the same repository.
All three tools offer business privacy controls, but terms depend on the product and plan:
Always verify the current privacy policy for your tier. Free tiers may have different terms.
Open the AI chat panel (Cmd/Ctrl+I), then click the mode selector dropdown at the top. Choose “Agent” for autonomous file editing or “Ask” for read-only exploration. You can also press Cmd/Ctrl+. to toggle modes.
Settings > Models > API Keys. Enter your Anthropic, OpenAI, or Google key, click Verify, then select “Use my API key” in the model dropdown. Own API keys work for chat models but not Tab completion or specialized features.
Cursor no longer documents the legacy Fast/Slow request buckets as the current plan model. Agent usage is metered against the allowance and rate card for your plan and selected model. Check the live breakdown in Settings > Usage; model price, Max Mode, and long contexts can consume included usage at different rates.
.cursor/bugbot.yml in your repositorycurl -fsSL https://claude.ai/install.sh | bashclaude --versionbrew install --cask claude-codeclaude --versionThen run claude in any project directory. You will be prompted to authenticate on first run.
The old progressive token-budget ladder is gone. think, think hard, and think more are ordinary prompt text. ultrathink is the one current exception: Claude Code recognizes it anywhere in a prompt and adds a deeper-reasoning instruction for that single turn, but it does not change the API effort level or reserve a fixed token budget. Use the thinking toggle and /model or /effort for durable control.
Granular approach (recommended): Configure allowed tools in your settings or use Claude Code’s permission system to pre-approve specific operations.
Full trust approach (use with caution):
claude --dangerously-skip-permissionsBest practice: Use --dangerously-skip-permissions only in sandboxed environments (Docker containers, CI pipelines).
Claude Code limits depend on plan, selected model, effort, context size, and tool use; Max plans raise included usage but are not unlimited. Long conversations with big files consume capacity faster. Check the live usage view rather than estimating a fixed message count. Strategies:
/compact to summarize long conversations/clear between unrelated tasksYes. Install the “Claude Code” extension from the VS Code marketplace for an integrated agent experience with editor context, diffs, and the same underlying Claude Code session — not merely an embedded terminal.
Codex is available across four surfaces:
codex). Interactive and headless modes. Open-source.Worktrees are an isolation option for local Codex tasks in the ChatGPT desktop app, not a universal property of every surface. When you choose an isolated desktop task, Codex uses a separate checkout so:
CLI sessions normally operate in the working directory you launch them from, while Cloud uses its own remote environment. For an isolated desktop task, Codex manages the worktree lifecycle.
Codex can respond to external events automatically:
@codex or mention it in a PR commentEach automation spawns a cloud agent that executes the task and creates a PR or posts results back to the source.
Approval policy and sandboxing are separate controls. Use on-request for interactive work so Codex can request an elevated action when needed. never means it will not ask: failures remain failures unless the configured sandbox already permits the operation. Pair never only with a deliberately chosen sandbox for trusted unattended jobs. The old --full-auto flag is a deprecated compatibility shortcut; configure approval and sandbox explicitly instead.
MCP (Model Context Protocol) servers extend what your AI agent can do by connecting it to external tools — databases, APIs, browsers, and more. You do not strictly need them, but they unlock powerful workflows. For example, the GitHub MCP lets the agent create PRs directly, and the Postgres MCP lets it query your database.
All three tools support MCP. Start with one or two essential servers and add more as needed.
MCP servers maintain a persistent connection and expose multiple tools. They are heavier to set up but more powerful for deep integrations (databases, browser automation, cloud services).
Agent Skills are lightweight, single-purpose augmentations installed via npx skills add <owner/repo>. They work across 35+ agents and are easier to share. Think of Skills as focused recipes and MCP servers as full integration platforms.
When both exist for a workflow, Skills are faster to set up; MCP servers offer deeper, persistent access.
Settings > Features > MCP > ”+ Add MCP Server”. Choose from the gallery or add a custom server configuration.
claude mcp add github -s user -- npx -y @modelcontextprotocol/server-githubOr add directly to .mcp.json in your project root.
codex mcp add github -- npx -y @modelcontextprotocol/server-githubOr add a [mcp_servers.github] table to ~/.codex/config.toml (project-scoped .codex/config.toml for trusted projects). The Codex IDE extension exposes this via MCP settings > Open config.toml.
A systematic approach for building complex features with AI:
/effort or use the thinking toggle for complex plans; think hard is ordinary prompt textThis works identically across all three tools. The key is giving the AI a structured starting point rather than a vague request.
Use Agent mode (or equivalent autonomous mode) when:
Use Ask/Chat/Suggest mode when:
/clear in Claude Code, new chat in Cursor/cost, Codex dashboard