Claude Code vs Codex
Claude Code and Codex are both agentic coding tools, and both run headless. They differ in reach and in control. Codex spreads across a desktop app, CLI, IDE extension and cloud, with native GitHub, Slack and Linear integrations. Claude Code concentrates on the terminal loop and gives more control over it: a full hook system, background subagents, and Anthropic models only.
At a glance
| Feature | Claude Code | Codex |
|---|---|---|
| What it is | Terminal-first coding agent | Multi-surface coding agent (desktop app, CLI, IDE, cloud) |
| Primary interface | Terminal, IDE, desktop app and browser; also Slack, mobile handoff and headless/SDK | Desktop App, with CLI, IDE extension and Cloud as secondary surfaces |
| Tab completions | Not available | Good, via the IDE Extension |
| Agent capability | Terminal-native agent loop with background subagents and dynamic workflows | Local, worktree and cloud task execution, with worktree threads for parallel work |
| Approval modes | Manual (default), auto, plan and bypass where policy permits | Auto, Read-only and Full Access |
| Context window | 1M tokens on current Sonnet, Opus and Fable tiers | Model- and surface-dependent; GPT-5.6 API: 1.05M |
| Models | Claude Fable 5, Opus 5 and Sonnet 5, switchable mid-session with `/model`; BYOK via API key | GPT-5.6 Sol, Terra and Luna, by plan and workload; BYOK via API key |
| MCP servers | Yes, via `.mcp.json` or `claude mcp add` | Yes, via `~/.codex/config.toml`, shared across surfaces |
| Agent Skills | Yes, via `npx skills add <owner/repo>` | Yes, via `npx skills add <owner/repo>` |
| Headless / scripted | `claude -p "prompt"`, plus web and managed agents | `codex exec "prompt"` |
Which one integrates better with GitHub?
Codex, if you want the review to happen in GitHub itself: PR reviews and `@Codex` are native, and Slack and Linear are too. Claude Code reaches the same outcome through GitHub Actions, web and desktop agents, and its own `/review` — more wiring, more control over what runs.
How do their safety controls differ?
Codex exposes three approval modes — Auto, Read-only and Full Access — plus hooks and approvals. Claude Code exposes four — manual, auto, plan and bypass where policy permits — and a full hook system that fires before and after every tool use, which is the difference between asking the agent to behave and making a command impossible to run.
Which models can each one run?
Neither crosses vendors. Claude Code runs Claude Fable 5, Opus 5 and Sonnet 5; Codex runs GPT-5.6 Sol, Terra and Luna. Both accept your own API key. If model choice is the deciding factor rather than workflow, Cursor is the tool that runs models from several vendors.
The verdict
Choose Claude Code if
- You want deterministic guardrails through hooks, not just approval prompts
- Background subagents and dynamic workflows match how you decompose work
- You are already on a Claude plan, which includes Claude Code
Choose Codex if
- You want work to start from GitHub, Slack or Linear without a terminal
- You want one agent across a desktop app, CLI, IDE and cloud
- You are already on a ChatGPT plan, which includes Codex
What they cost
Claude Code
- Claude Pro
- Claude Code included; plan usage is shared with Claude
- Claude Max 5x
- Five times Pro session capacity; Claude Code included
- Claude Max 20x
- Twenty times Pro session capacity; Claude Code included
Teams: Team Standard $25/member/month ($20 billed annually); Team Premium $125/member/month ($100 annually); Enterprise is custom
Codex
- ChatGPT Plus
- Codex included; credits may extend usage after included limits
- ChatGPT Pro 5x
- Same Pro capabilities with five times Plus usage
- ChatGPT Pro 20x
- Highest individual allowance; twenty times Plus usage
Teams: ChatGPT Business seat $25/user/month ($20 billed annually); Enterprise is custom
Public US list prices before tax, from the snapshot date above. Regional prices, negotiated contracts, mobile-store billing and promotions can differ — check the vendor page before purchasing.
Frequently asked questions
Is Codex included in a ChatGPT plan?
Yes. Codex is included in eligible ChatGPT plans rather than sold as a separate consumer subscription, starting with ChatGPT Plus at $20/month. Credits may extend usage once the included limits are reached.
Can both run without a human present?
Yes. Claude Code runs claude -p "prompt" plus web and managed agents; Codex runs codex exec "prompt". Both emit structured JSON output, so both fit into CI. Both also support scheduled automations.
Which has better goal-directed runs?
Both have /goal. Claude Code runs a separate evaluator model each turn; in Codex /goal has been stable and on by default since CLI 0.133.0 and persists across sessions. Codex has no recurring-prompt equivalent in CLI 0.145, where Claude Code bundles /loop.