codex exec on a ChatGPT login: measured overhead, schema pollution and config traps

codex exec on a ChatGPT login

Tested 14 October 2026 with codex-cli 0.162.1 on Windows, signed in with ChatGPT (codex login status reports "Logged in using ChatGPT").

Setup notes

• codex was installed through npm, under %APPDATA%\npm.
• With an OPENAI_API_KEY environment variable also present, codex login status still reported "Logged in using ChatGPT", so Codex kept using the subscription rather than the API key.
• ~/.codex/config.toml is shared with the Codex desktop app. In the observed setup it set sandbox_mode = "danger-full-access" and had no forced_login_method. Passing -s read-only or -s workspace-write explicitly, or --ignore-user-config, avoids inheriting that.

Gotchas

• Fixed overhead is about 38k input tokens per call, of which about 17k was cached. This was measured on a one-line prompt with --ignore-user-config, so plugins and MCP servers were not loaded. It is higher than the "about 22k" figure that circulates, which makes the tool a poor fit for many small calls.
• --output-schema checks the shape but not the values. Codex ran a web search unprompted and put a markdown citation inside a string field, for example "capital":"Wellington (city population source)". Instruct it not to search or cite, or validate field values.
• In Python, subprocess.run(["codex", ...]) does not find the codex.CMD npm shim. Use shutil.which("codex") to get the full path, or pass shell=True.
• The Codex CLI cannot run inside a Cloudflare Worker, so high-volume small-call workloads there need a hosted model API instead.

Working command

echo "<prompt>" | codex exec --skip-git-repo-check --ephemeral --ignore-user-config -C <dir> -s read-only -c model_reasoning_effort="low" --output-schema schema.json -o last.txt --json -

The --json event stream includes turn.completed with usage (input_tokens, cached_input_tokens, output_tokens).