Skip to content
The Test Cabinet The Test Cabinet

Overview

OpenAI Codex (slug codex) is OpenAI’s coding agent CLI, driven non interactively through its codex exec subcommand. It authenticates against OpenAI directly with an API key and runs the model OpenAI serves natively, so a Codex run reports the vendor’s own model IDs and usage figures. See the OpenAI Codex site for the harness itself.

Model IDs

Section titled “Model IDs”

Codex uses OpenAI’s vendor-native model names unchanged — there is no provider prefix or routing slug. Examples (illustrative, not exhaustive):

  • gpt-5.5
  • gpt-5.4-mini
  • gpt-5.3-codex-spark

Invocation

Section titled “Invocation”

The harness probes and invokes the codex binary. It is installed into the run container at run time with:

Terminal window
npm install -g @openai/codex && npm cache clean --force

A session is run with codex exec in non-interactive mode. The Test Cabinet passes these flags:

FlagPurpose
--jsonEmit the line-delimited JSON event stream consumed for events and usage.
--skip-git-repo-checkRun outside a git repository without prompting.
--dangerously-bypass-approvals-and-sandboxRun unattended, without per-action approval prompts.
--model <id>The model to run.

The prompt is passed as the final positional argument.

Authentication. Codex authenticates with either an OpenAI API key or a ChatGPT account subscription; by default a subscription is preferred when you are signed in. With the API key, the key is sourced from OPENAI_API_KEY on the host and — because codex exec reads its key only from CODEX_API_KEY and ignores OPENAI_API_KEY — injected into the run container under the name CODEX_API_KEY. See Authentication for both modes and how to lock one.

Pricing. Codex reports bare OpenAI model IDs, so the comparable-cost lookup prepends an openai/ prefix before consulting OpenRouter — gpt-5.5 becomes openai/gpt-5.5. Codex does not self-report a run cost, so the comparable cost is always OpenRouter-derived; see Metrics.

See Authentication for the API-key and subscription modes, Events for how Codex’s output is normalized, and Metrics for how its usage is counted. For the harness layer these pages fit into, see Harnesses.