OpenAI Codex: Current Setup and Models

OpenAI Codex is a coding agent available through the CLI, IDE extension, app, and cloud workflows. The current official recommendation is GPT-5.5 for demanding Codex tasks and GPT-5.4 mini for faster, lower-cost work.

GPT-5.2 and GPT-5.3-Codex are deprecated for ChatGPT-authenticated Codex sessions. GPT-5.6 Sol, Terra, and Luna are newer, but they are available only to selected approved API organizations and Codex workspaces during the preview.

Install

1
2
3

npm install -g @openai/codex
codex --version
codex

Codex can authenticate with ChatGPT or an API key:

• ChatGPT authentication includes Codex according to the user’s Free, Go, Plus, Pro, Business, Edu, or Enterprise plan.
• API-key authentication uses models available to that API key and API token billing.
• API-key authentication does not provide Codex cloud features such as hosted code review or Slack integration.

Do not hard-code an expected CLI version in an evergreen setup guide. Verify the installed version and current release notes.

Current Model Selection

Set the local default in ~/.codex/config.toml:

1

model = "gpt-5.5"

Choose temporarily:

1

codex -m gpt-5.5

Or use /model in the CLI and the model selector in the IDE extension. Codex cloud tasks currently do not expose a user-selectable default model.

Use the GPT-5.6 guide for Sol/Terra/Luna pricing and the Sol investigation for agent-risk controls.

AGENTS.md

AGENTS.md is Markdown guidance, not an invented YAML agent registry. Put repository conventions, commands, boundaries, and verification requirements in the repository root:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16

# Repository guidance

## Commands
- Install: `bun install`
- Test: `bun test`
- Build: `bun run build`

## Change rules
- Keep patches scoped to the requested feature.
- Do not add dependencies without approval.
- Never commit secrets or generated credentials.

## Verification
- Run the focused test first.
- Run the build before handoff.
- Report any skipped checks and why.

Closer nested AGENTS.md files can provide subtree-specific instructions. Model, provider, approval, and sandbox defaults belong in Codex configuration, while task-specific constraints belong in the prompt.

Approvals and Sandbox

Start with the narrowest permissions that allow the task:

• keep production credentials out of the agent environment;
• require approval for network access, destructive commands, or writes outside the workspace;
• use clean worktrees for isolated feature work;
• review the diff and tests instead of trusting the final summary;
• record any escalated command and its purpose.

These controls are especially important for long-running agents. OpenAI’s GPT-5.6 preview system card documents simulated cases of destructive scope expansion, credential movement, and unverified completion claims.

Local and Cloud Work

Local Codex surfaces work with the checked-out repository and the configured sandbox. Cloud tasks use a configured cloud environment and hosted execution. Check environment setup, secrets, internet access, and repository state separately for each surface.

Do not promise fixed concurrency, microVM startup time, message counts, or credit consumption unless the current official plan or account UI documents it. Limits vary by plan, task size, model, and rollout.

Plan Overview

The official June 28 pricing page lists:

Plans, regional pricing, credits, and limits can change. Use the official pricing page and signed-in account as final authority.

Current Model Evidence

Benchmarks do not evaluate the entire Codex product. Measure accepted patches, tests, retries, wall-clock time, and review effort in the actual sandbox and repository.

Migration From Old Configurations

Search configuration and scripts for deprecated models:

1

rg 'gpt-5\.2|gpt-5\.3-codex' ~/.codex . --glob '*.toml' --glob '*.json' --glob '*.md'

For ChatGPT-authenticated Codex, replace active defaults with gpt-5.5. Leave historical articles, fixed eval baselines, and API-specific integrations unchanged unless their runtime model is also being migrated.

Sources

• OpenAI Codex manual: Overview, models (Archive), pricing (Archive), and configuration
• OpenAI: GPT-5.6 preview access
• OpenAI: GPT-5.6 Preview system card

Related links

• /models/gpt-5-6/
• /compare/codex-vs-claude-vs-kimi/
• /compare/codex-vs-claude-vs-cursor/
• /risks/codex/cloud-dependency-risks/
• /agents-md/

Last verified: June 28, 2026 against the current official Codex manual. Plans, models, defaults, preview access, limits, and surface behavior can change independently.