Independent analysis of AI coding plans

What each subscription costs, and the most it could possibly hand you at the vendor's own API rates.

By the numbers

How this works

Step one

Take the vendor's own price list

Every model here has a published per-token API rate. That rate is what the vendor charges anyone off the street, so it is the fairest possible yardstick for what a token from them is worth.

Step two

Find out what the plan lets through

This is the hard part, because almost nobody publishes it. A plan is metered in prompts, calls, messages, credits or an unnamed multiplier, and none of those convert to tokens without measuring.

Step three

Multiply, then divide by the price

Tokens the plan allows, times the vendor's own rate, over what you pay. That is the subsidy. Where step two is blank the answer is blank too, and no estimate is put in its place.

Highlights

Cost of a million tokens

Blended list API rate, input weighted 3:1 over output · Lower is better

Plans

Plan Plan price 5h limit Weekly limit Max possible spend (approx.)

Max possible spend is a ceiling, not a forecast: it assumes every window of the month runs flat out, which is nobody's real life. No vendor publishes it. It is only computable where somebody has measured a saturated window and the vendor publishes a multiplier to scale it across tiers. Everything else reads not measured, and that blank is the finding.

Measure your own
node probe/probe.js --days 30 --plan claude-max-20x
node probe/windows.js --days 30

The first prices what you burned. The second buckets every turn into its rolling window and reports the distribution, which is what turns a price into a ceiling. Both read the token counts Claude Code and Codex already write to ~/.claude/projects and ~/.codex/sessions. Nothing leaves your machine unless you pass --submit.