As of November 2026, pick Codex if you have a ChatGPT plan and want OpenAI's own agent with an operating-system sandbox and one login. Pick OpenCode if you want an open source (MIT) agent where you choose the provider and model, including local models and a list of free ones. Both read the same AGENTS.md, so many developers install both and run them on different tasks.
Everything below was checked on 19 November 2026 against the official OpenCode docs, which now document OpenCode 2, and OpenAI's Codex CLI docs. If you are also weighing Anthropic's agent, see OpenCode vs Claude Code and Codex vs Claude Code.
OpenCode vs Codex at a glance
| OpenCode | Codex | |
|---|---|---|
| Maker | The OpenCode team (anomalyco/opencode on GitHub) | OpenAI (openai/codex) |
| Licence | Open source, MIT | CLI is open source, Apache-2.0 |
| Models | Any connected provider from the models.dev catalog, local servers, OpenCode Console models | OpenAI models, such as GPT-6.2 Astra and GPT-6.2 Sol; --oss for a local provider |
| How you pay | Provider API keys, ChatGPT Plus or Pro, GitHub Copilot, pay-as-you-go Console credits, Go subscription, some free models | ChatGPT plan (Codex is included), or an API key at API pricing |
| Platforms and install | macOS, Linux, Windows (standalone zip); curl script, npm, Homebrew | macOS, Linux, Windows; install script, npm, Homebrew |
| Interfaces | Terminal UI, desktop app, web UI, mobile app | CLI, IDE extension, ChatGPT desktop app, Codex web |
| Approvals and sandbox | Ordered permissions rules (allow, ask, deny); no sandbox described | Sandbox modes plus approval policies |
| Instructions file | AGENTS.md | AGENTS.md, created with /init |
| Session resume | /sessions, opencode run --continue | codex resume |
| MCP | Yes, opencode mcp add | Yes, codex mcp |
| Headless mode | opencode run, opencode serve | codex exec |
| Config | opencode.json | config.toml |
Where does the model come from?
Codex runs OpenAI models on your ChatGPT or API account, and OpenCode runs whichever provider you connect.
In Codex you start codex in a project folder, choose "Sign in with ChatGPT" or another sign-in method, and pick a model and reasoning effort with /model. As of November 2026 the Codex models page lists GPT-6.2 Astra for demanding work and GPT-6.2 Sol for everyday coding, which we cover in GPT-6.2 Astra and Sol. You can also pass a model directly, for example codex -m gpt-6.2-sol. The --oss flag points Codex at a local open source model provider. Our guides to installing the Codex CLI and Codex models cover setup and defaults.
OpenCode ships a provider and model catalog from models.dev. You run /connect (or opencode auth login outside the interface), pick a provider and add credentials, then choose with /models. opencode models lists everything available. The providers page says Ollama, LM Studio and vLLM servers are discovered locally. Our OpenCode models guide shows defaults and variants.
Is OpenCode free? Is Codex?
Both agents cost nothing to install, and both have a way to run at no model cost, with limits.
For OpenCode, as of November 2026, the Console models page (the former Zen page) lists free models such as Big Pickle and DeepSeek V4.1 Flash Free, each described as free for a limited time and some with data retention caveats. Paid Console models are pay as you go, with a processing fee of 4.6% plus $0.31 on each credit purchase. OpenCode Go is an optional subscription at $10 a month. Our list of OpenCode free models tracks what is free now.
For Codex, OpenAI's pricing page says, as of November 2026, that Codex is included in ChatGPT Free, Go, Plus, Pro, Business, Edu and Enterprise plans. It lists Go at $8 a month, Plus at $20 and Pro from $100. Each plan has its own usage allowance. With an API key, usage is billed at API pricing, separate from your plan.
Can I use my ChatGPT subscription in OpenCode?
Yes, according to OpenCode's own docs.
The OpenCode providers page says you can use a ChatGPT Plus or Pro plan in place of an OpenAI API key: connect OpenAI, choose "Sign in with ChatGPT", and usage counts against your plan. GitHub Copilot connects through device sign-in. The OpenAI pages we checked describe signing in to Codex itself and do not mention OpenCode, so treat the OpenCode docs as the source for this and recheck after upgrades.
That means one ChatGPT plan can drive both tools. What you get from Codex in return is OpenAI's sandbox, its IDE extension, the ChatGPT desktop app and Codex web.
How do sandboxing and approvals differ?
Codex sandboxes commands at the operating-system level, while OpenCode decides per action with permission rules and describes no sandbox.
Codex separates two ideas. The sandbox mode sets what the agent can technically do: read-only, workspace-write or danger-full-access. The approval policy sets when it must ask: on-request, never or a granular table. The approvals and security docs say the older untrusted policy is retired. In a version-controlled folder Codex defaults to Auto, which is workspace-write with on-request approvals, and network access is off by default. Folders that are not under version control start read-only. The sandbox uses Seatbelt on macOS and bwrap with seccomp on Linux. Set these with --sandbox and --ask-for-approval, or with sandbox_mode and approval_policy in config.toml. Our Codex sandbox guide has examples.
OpenCode 2 uses an ordered permissions array of action, resource and effect rules, where the last matching rule wins. Its permissions page says the base policy allows every action and only asks before reading .env files, and that shell commands run with your user's filesystem, process and network authority. The Plan agent cannot edit normal project files. Press Shift+Tab to cycle agents. To be asked before each shell command:
{
"permissions": [
{ "action": "shell", "resource": "*", "effect": "ask" }
]
}OpenCode 1 named the shell action bash. Version 2 still reads old configs. Neither approach replaces review: a sandbox limits damage, it does not tell you whether the change is correct.
Which is better for large refactors and long tasks?
Neither wins on a measurement we can cite, so compare the process. We publish no benchmark here.
- Start read-only. Use OpenCode's Plan agent or Codex's
read-onlysandbox, read the proposal, then let it edit. - Isolation. Codex has a
--worktreeflag that starts a session in a new managed Git worktree (see Codex worktree). With OpenCode you create the worktree yourself withgit worktree add. - Undo. OpenCode has
/undoand/redo. In Codex, the workspace sandbox and Git are your safety net. - Away from the keyboard. Codex web runs tasks in the cloud, and
opencode serveruns OpenCode as a server you can reconnect to. - Limits. Codex draws from your ChatGPT plan allowance or API billing. In OpenCode the limit is whatever the connected provider sets, and you can switch models with
/models.
How do you script each one?
Use codex exec for Codex and opencode run for OpenCode.
codex exec (alias codex e) runs non-interactively. The command reference lists --json for newline-delimited events, --output-last-message (-o) to write the final message to a file, and --output-schema for a JSON Schema. Our codex exec guide goes further.
opencode run takes --model, --agent, --continue, --session and --format json, per the OpenCode command reference. opencode serve starts the API and web server.
# Codex: one scripted task, final message saved to a file
codex exec --sandbox workspace-write -o result.md "Add tests for the date parser"
# OpenCode: one scripted task with an explicit model
opencode run --model openai/gpt-6.1-sol "Add tests for the date parser"Can you use both together?
Yes. They share AGENTS.md, keep separate sessions, and can work on the same repository if each task owns its files.
Keep build commands, layout and conventions in AGENTS.md, and tool-specific settings in config.toml and opencode.json. Our guide to AGENTS.md for Codex explains discovery and size limits. Two agents editing the same files will collide, whichever tools they are, so give each parallel task its own Git worktree.
You do not have to choose at all if you coordinate them. VibeiDE is a desktop app that coordinates the Codex, Claude Code and OpenCode CLIs you already installed, with your own provider accounts. Codex runs in an embedded terminal that shows the original CLI, while OpenCode runs headless with recorded activity. Each project has its own task queue, so a Codex task and an OpenCode task can run on the same project, and the "New worktree per task" mode gives each its own worktree and branch. Finished tasks wait in Ready for review. See the Codex page and the OpenCode page, or read about running agents in parallel and the plan council.
How do you switch between them?
Install the other tool and sign in. The AGENTS.md file carries over unchanged.
From Codex to OpenCode:
curl -fsSL https://opencode.ai/v2/install | bash
cd your-project
opencodeThen run /connect, choose OpenAI and "Sign in with ChatGPT" to keep the same plan, or pick another provider. Our OpenCode install guide covers Windows and the desktop app.
From OpenCode to Codex:
curl -fsSL https://chatgpt.com/codex/install.sh | sh
cd your-project
codexCodex is also on npm (npm install -g @openai/codex) and Homebrew (brew install --cask codex). MCP servers are configured per tool, with opencode mcp add and codex mcp. Sessions do not transfer. If neither fits, see our list of OpenCode alternatives.
FAQ
Is OpenCode a replacement for Codex?
No. They overlap, but they are separate tools. OpenCode lets you choose providers, and Codex is OpenAI's agent. Many developers use each where it fits.
Does OpenCode work with ChatGPT Plus?
Yes. OpenCode's docs describe a "Sign in with ChatGPT" option for Plus and Pro plans, with usage counted against the plan.
Is Codex CLI open source?
Yes. The openai/codex repository is licensed under Apache-2.0. OpenCode's repository is MIT licensed.
Which is safer, OpenCode or Codex?
Codex's default in a version-controlled folder limits writes to the workspace and keeps network access off. OpenCode 2 allows actions by default unless you add permissions rules. Review the diff either way.
Takeaways
- Codex is OpenAI's agent with sandbox modes and approval policies. OpenCode is MIT licensed, provider-agnostic and uses
permissionsrules. - As of November 2026, Codex is included in ChatGPT plans, and OpenCode documents a ChatGPT Plus or Pro sign-in.
- Both read
AGENTS.md, so one shared file works. - Use
codex exec, oropencode runandopencode serve, for scripting. - Running both on one repository needs one task per session and a worktree per parallel task.





