[bin] Explore creating a CLI helper for parsing OpenAI codex token usage and computing costs

[0xdevalias]

Explore creating a CLI helper that can parse the OpenAI Codex token-usage output format and compute detailed cost breakdowns. The primary use case is taking a typical Codex usage string such as:

Token usage: total=112,632 input=67,873 (+ 1,148,032 cached) output=44,759 (reasoning 37,312)

Extracting the relevant token counts (input, cached, output, reasoning, etc.), looking up model pricing, and calculating & displaying the cost for each section and the total.

The tool should also support being invoked with individual token parameters (e.g. --input=… --output=… --reasoning=…) for situations where the usage string isn’t available or when doing custom cost estimation.

Model & Pricing Sources / Constraints

We need solid sources to get model/pricing info. Some options & caveats:

• We can lookup pricing on the OpenAI API pricing page:
  • https://openai.com/api/pricing/
• Also, Codex-specific pricing info appears on the OpenAI Codex pricing page:
  • https://developers.openai.com/codex/pricing/
• Using the 'Using Codex with your ChatGPT plan' doc helps understand which subscription tiers include Codex and what usage limits apply:
  • https://help.openai.com/en/articles/11369540-using-codex-with-your-chatgpt-plan
• Currently there is no public API endpoint from OpenAI that returns model pricing programmatically.
  • Developers have confirmed this in forums.
  • There’s a feature request in openai-python asking for this.
• Third-party/community maintained sources may have helpful data, but may lag or be inaccurate; these should be used with caution.

Proposal: Implementation Notes

• Maintain a pricing config (JSON/YAML) mapping relevant models (Codex, codex-mini, etc.) → cost per token category (input, output, cached, reasoning). Include metadata like “last verified” date.
• Support two input modes:
  1. Full usage string in Codex format (as above)
  2. Individual token flags (e.g. --input, --cached, --output, --reasoning) for custom or partial data.
• Handle unknown/new models gracefully (warn user, allow manual override of pricing, fallback to a default model or default rates).
• Optionally a command or flag like --update-pricing which attempts to fetch/prune pricing from official sources (scraping or via OpenAI API if/when one exists). Robust handling for errors or stale data.
• If relevant, support currency options or conversion (e.g. default USD, maybe allow AUD etc.) to make it usable locally.

Example CLI Output Mockup

$ bin/codex-token-cost codex-mini "Token usage: total=112,632 input=67,873 (+ 1,148,032 cached) output=44,759 (reasoning 37,312)"

Model: codex-mini
Pricing: input $1.50 / 1M tokens, output $6.00 / 1M tokens, cached $0.75 / 1M tokens, reasoning $X / 1M tokens

  • Input tokens:         67,873        → $0.102
  • Cached tokens:      1,148,032        → $0.861
  • Output tokens:         44,759        → $0.269
  • Reasoning tokens:      37,312        → $0.XXX
  ─────────────────────────────────────────────

[0xdevalias]

I added an initial version of this (called codex-session-cost) a couple of weeks back in 5306196:

• https://github.com/0xdevalias/dotfiles/blob/5306196b8d8fc878e0595f96be3ff0b69fd3a474/bin/codex-session-cost

I had forgotten I had an issue open about it, so it may not fully align with the notes / plan laid out here previously.

[Necmttn]

I would keep pricing provenance as visible as the math. For every estimate: parsed token buckets, parser pattern version, model id, model_source, pricing source URL/file, last_verified date, and whether reasoning/cached rates were known or manually supplied.

That avoids the dangerous failure mode for cost helpers: a new Codex usage string parses successfully, but the model/pricing assumption is stale and the total looks authoritative.

Generated with ax - https://github.com/Necmttn/ax