OpenAI Codex: Current Setup and Models

OpenAI Codex is a coding agent available through the CLI, IDE extension, app, and cloud workflows. GPT-5.6 Sol, Terra, and Luna have been generally available in Codex since July 9.

GPT-5.2 and GPT-5.3-Codex are deprecated for ChatGPT-authenticated Codex sessions. Choose the GPT-5.6 tier and effort level that fit the task rather than treating GPT-5.5 as a required baseline.

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-5.6-terra"

Choose temporarily:

1

codex -m gpt-5.6-sol

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 current ChatGPT, Codex, and API access, pricing, and system-card status. The Sol preview investigation is retained as June history.

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 July 9 GPT-5.6 system card reports low-frequency simulated cases of scope expansion beyond user intent and describes confirmation and safety controls.

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.

Usage limits and resets

Current OpenAI pricing documents shared five-hour windows and possible weekly limits, while current account reports show weekly-only displays. No first-party deprecation or completed-restoration notice resolves the difference. Model choice, reasoning effort, context, tool calls, searches, caching, retries, and subagents can change how quickly a task consumes the active allowance.

Check the signed-in usage dashboard and /status before purchasing overflow. Some historical or supported clients also expose /usage. Keep these paths separate:

• scheduled recovery ends the current allowance window;
• provider-wide or incident resets refresh the named group;
• banked resets are expiring, eligible user-controlled grants;
• purchased full resets are an account-scoped rollout reported at $8 on some Plus accounts and $80 on some Pro 20x accounts; redemption is reported to start a new seven-day window rather than stack a bonus week; and
• purchased credits fund supported work after included limits without a documented counter refill.

The paid-reset prices and availability are community-observed, not universal OpenAI terms. Pro 5x, Business, eligibility, refund behavior, and future prices remain unresolved.

Read AI Subscription Capacity Is Perishable for the scheduling rule, the reset chronology for dated scope, and the cost-saving playbook before choosing between waiting, a displayed reset, credits, or another route.

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 deprecated defaults with a GPT-5.6 tier available to the account and suited to the workload. Leave historical articles and fixed eval baselines unchanged.

Sources

• OpenAI Codex manual: Overview, models (Archive), pricing (Archive), and configuration
• OpenAI: GPT-5.6 general availability
• OpenAI: GPT-5.6 System Card
• Community paid-reset observations: $8 Plus, $80 Pro 20x and shifted recovery, and uneven rollout

Related links

• /models/gpt-5-6/
• /posts/prompt-caching-agent-token-costs/
• /posts/ai-usage-limits-use-it-or-lose-it/
• /posts/codex-banked-resets-gpt-5-6-preview/
• /posts/openai-exploitgym-hugging-face-breach/
• /compare/codex-vs-claude-vs-kimi/
• /compare/codex-vs-claude-vs-cursor/
• /risks/codex/cloud-dependency-risks/
• /agents-md/
• /value/llm-cost-saving-playbook/

Last verified: August 20, 2026. Plans, models, defaults, limits, prices, and account behavior can change independently.