Real API Pricing divides each coding-agent subscription fee by the monthly tokens the plan lets a customer spend, across 184 plans

FeiZhuLulu's Real API Pricing repository, published on GitHub on 5 September 2026, computes one unit price for every AI coding subscription and metered API on its list: the monthly subscription fee divided by the tokens the plan lets a customer spend in a month. The project's 2026-09-07 snapshot covers 184 plan-and-model rows, 173 of them subscriptions with a monthly allowance and 11 of them metered API baselines. It has picked up 130 stars in the three days since launch, and ships under MIT.

The workload it assumes, and says so on the front page

The number is only fair if everyone is measured against the same imagined usage. The repository states its convention in one sentence at the top of the README: 97.5 percent cache reads, 2.15 percent fresh input, and 0.35 percent output. That mix converts a dollar-and-credit pool or a three-part token price into one comparable dollar-per-million-tokens figure. Where a plan already publishes a total-token dashboard, a local usage log, or an official absolute-token table, the observed number is used as is and not renormalised.

Where the numbers come from, plan by plan

GPT-5.6 Luna on ChatGPT Plus is derived from a dashboard measurement of 112.67 million total tokens consuming about 6 percent of the weekly allowance, which the maintainer reports as 7.511 billion tokens per month for Plus. The 5x and 20x plans are scaled off that Plus baseline, giving the top Luna point at 150.222 billion tokens per month, at medium confidence. The 240.24 billion Sol-credit estimate previously used is marked superseded and left in the audit trail rather than deleted.

Claude Max is quoted at a permanent 15.7 billion tokens per month under the terms that come into effect on 14 September 2026. GLM Coding Plan is recomputed from Zhipu's official weekly credits, split into peak, midpoint and off-peak scenarios instead of copying the 95-percent-cache example table from the official docs. Every conversion has a dated JSON file behind it under data/research/.

Which leaderboard's Y-axis, and why it matters here

The scoreboards this project draws its axes from (Code Arena, Agent Arena, AA Intelligence, AA Coding Agent) also come with rules of their own. The README repeats them so a reader is prepared for how each is scoped: Code Arena here refers specifically to the WebDev Overall Arena Score. AA Coding Agent points are archived highest-configuration reference summaries of a harness-model-effort setup, and each channel's own measured quota still has to be verified against real usage. When a provider only reports total tokens with an unknown input-and-output split, that limitation is written into the row rather than filled in with a guess.

The interesting comparisons come out at the ends. The dashboard's own Pareto charts show that once a subscription caps at a modest allowance, its real dollar-per-million-tokens figure moves the dot to the left on a log-scale price axis, well past the raw published sticker price on the same model bought through the metered API. Cache-heavy workloads make that gap wider, and providers that meter cache writes separately widen it again.

For a team choosing which coding-agent plan to buy, or which one to renew, the useful move is to open the repository's data/adopted.csv, read the row for the current plan, and compare it against the team's own workload. When the team's cache-hit rate sits well below 97.5 percent, the ranking on the bill will not match the ranking on the chart, and the repository writes down enough of its working that a lead can recompute it against real usage numbers.

Source

• FeiZhuLulu/real-api-pricing on GitHub